Main

Mycobacterium tuberculosis (Mtb), the world’s leading infectious cause of death1, is among the bacteria that have been most extensively sequenced, but the functions and interactions of many of its genes remain poorly understood. This knowledge gap limits our understanding of the genetic foundations of tuberculosis (TB) and hinders the development of effective therapies.

Genetic interaction (GI) mapping is a powerful approach for deciphering the functional relationships between genes2,3. It involves systematically assessing how perturbing one gene modifies the phenotype associated with perturbing a second gene. A GI is defined as a combined phenotype that deviates from the expectation based on single-gene perturbations. Positive interactions occur when dual perturbation produces a milder phenotype than expected, whereas negative interactions—including synthetic lethal effects—result in a more severe phenotype. Applied at scale, GI mapping yields quantitative interaction profiles that serve as high-resolution phenotypes3. These profiles can generate functional hypotheses for uncharacterized genes, reveal surprising roles for genes of known function and identify connections between distinct biological pathways3. Moreover, they can highlight potential synergistic and antagonistic drug–target relationships, informing combination therapies5.

GI mapping in Mtb has historically been constrained by biological and technical challenges, including slow growth, biosafety requirements and limited early genetic tools. The development of high-density transposon mutagenesis libraries combined with hybridization- or sequencing-based readouts (TnSeq)6,7 constituted a major advance, permitting genome-wide identification of conditionally essential genes8 and GI mapping in deletion mutants9. However, such approaches are typically restricted to non-essential genes, excluding most antibiotic targets. CRISPR interference (CRISPRi) has since emerged as a complementary technique10,11,12,13, enabling conditional and tunable knockdown of both essential and non-essential genes. In bacterial systems, dual-gene knockdowns have been used to identify GIs among small gene sets14, and larger-scale efforts have paired individual deletion or knockdown strains with genome-wide libraries15,16. Dual-CRISPRi and dual-TnSeq approaches for comprehensive GI mapping have been implemented in Bacillus subtilis and Streptococcus pneumoniae17,18,19. However, technical limitations have thus far precluded the development of a comparable system in Mtb.

To enable high-throughput GI mapping in Mtb, we developed a dual-single-guide RNA (sgRNA) CRISPRi platform for mycobacteria. Addressing key challenges in library construction and data analysis, we devised a robust experimental and statistical pipeline for high-confidence detection of GIs. Using this platform, we mapped GIs among nearly all essential genes in Mycobacterium smegmatis (Msm) and Mtb, prioritizing this technically demanding class of genes because of their centrality to cellular processes and importance as antibiotic targets. These maps recapitulate known biological relationships, validating the platform, and also uncover thousands of novel GIs. We further performed AlphaFold3 co-folding for all gene pairs (approximately 600,000) represented in the GI maps, identifying GIs that are likely to reflect physical associations and offering a basis for mechanistic hypotheses. Finally, we validated selected predictions from the maps, including drug–drug interactions and roles in the electron transport chain (ETC) for several genes of unknown function. Together, these maps constitute a powerful resource for elucidating Mtb biology and guiding the rational design of combination therapies.

Optimal sgRNAs for GI mapping

To enable dual-sgRNA CRISPRi-based GI mapping in Mtb, we adapted the Combinatorial Genetics En Masse (CombiGEM) framework20 to the Mtb CRISPRi platform10, assigning each sgRNA a barcode, and using a cloning workflow that preserves sgRNA–barcode linkage and positions paired barcodes in close proximity (Extended Data Fig. 1a). This design permits accurate quantification of dual-sgRNA strain abundance by barcode sequencing, facilitating high-throughput GI mapping.

GI detection for essential genes requires sgRNAs that impose a precise fitness cost: sufficient to reveal interactions, yet not so severe as to mask additive or synergistic effects. Compounding this challenge, the combinatorial scale of dual-sgRNA libraries necessitates selecting only a small number of sgRNAs per gene. To define the optimal fitness cost for GI detection, we performed a pilot screen in Msm using a panel of 16 genes that are predicted to interact. We selected up to 24 sgRNAs per gene (Supplementary Methods and Supplementary Table 1) spanning a broad range of fitness costs, cloned a dual-sgRNA library containing all pairwise permutations and performed a GI screen using the workflow planned for larger-scale studies (Extended Data Fig. 1b). The resulting data allowed us to identify the sgRNA strength window that best supports GI detection under these conditions: an approximately 8% reduction in fitness per generation (Extended Data Fig. 1c–h and Supplementary Discussion).

Essential-gene GI map in Msm

To establish feasibility, we built a comprehensive essential-gene GI library in Msm. Guided by the pilot screen, we selected up to two sgRNAs per gene with reproducible, intermediate fitness costs (Fig. 1a and Supplementary Table 1). The library contains 1,524 sgRNAs targeting 787 genes, plus 10 non-targeting (NT) controls, yielding about 2.3 million sgRNA permutations. Target genes included those that are essential for Msm fitness and those that are non-essential in Msm but essential in Mtb; the latter were targeted by the predicted strongest sgRNAs10. The library was cloned in Escherichia coli, validated by deep sequencing (Extended Data Fig. 2a) and transformed into an Msm strain containing a chromosomally encoded Streptococcus thermophilus (Sth1) dCas9 expression cassette. This yielded a pooled library in which each bacterium carries a single integrated guide plasmid expressing two sgRNAs. We subjected this library to competitive growth, tracking dual-sgRNA strain abundance over time by barcode sequencing. Strain depletion was modelled using a piecewise linear regression10, providing a quantitative readout of fitness.

Fig. 1: Comprehensive GI map of essential genes in Msm.

a, Schematic of the screen. (1) Up to two sgRNAs per gene (log2-transformed fold change (L2FC) –1 to –5) were selected. (2) sgRNAs were assigned barcodes (BCs), and all pairwise permutations were generated. (3) The library was passaged with or without anhydrotetracycline (ATc), and L2FC trajectories were computed from barcode counts. (4) GI scores were calculated for each gene pair (see e,f), and genes were clustered by GI profile correlation. (5) Applications of GI data. b, Distribution of observed L2FC at generation 20 for dual non-targeting (NT), single-targeting (A + NT) and dual-targeting (A + B) constructs. c, Correlation of observed L2FC at generation 20 for single-targeting constructs differing only in sgRNA position (Pos1, position 1; Pos2, position 2). d, Multiplicative model of GIs for sgRNAs targeting genes A and B. Fitness outcomes greater or less than expected indicate positive or negative GIs, respectively. e, Example of GI quantification. Left: L2FC trajectories for a representative dnaJ1 + dnaJ2 dual-sgRNA construct. Right: ΔL2FC (observed − expected) trajectories for all eight dnaJ1 + dnaJ2 constructs. f, Expected versus observed L2FC at generation 25 for all dual-targeting constructs. Black line, x = y; red line, generalized additive model (GAM) fit. GI scores were computed as the median GAM residual per gene pair. g, Distribution of GI scores across all gene pairs. Shaded regions denote statistically significant (p > 0.5) interactions. h, Heat map of GI scores for all gene pairs, hierarchically clustered by Pearson correlation of GI profiles. Values beyond ±5 are clipped. i, Number of positive and negative GIs (p > 0.5) per gene. j, Distribution of pairwise GI profile correlations among all sgRNAs, and among sgRNAs targeting the same gene. k, GI scores for all genes paired with murA (left) or with cwlM (right). l, Left, dendrogram of all genes, clustered by GI profile correlation. Branches are coloured by nodes enriched (false discovery rate (FDR) < 0.005) for Gene Ontology (GO) terms by Fisher’s exact test. Right, representative clusters.

Growth phenotypes were highly reproducible across biological replicates (Extended Data Fig. 2b,c). As expected, strains with two NT sgRNAs showed minimal depletion, single-targeting strains exhibited moderate depletion and dual-targeting strains showed the greatest depletion (Fig. 1b). sgRNA position within the plasmid had little effect on fitness, indicating consistent activity from both expression cassettes (Fig. 1c).

To identify GIs, we compared the observed fitness of each dual-targeting strain to the expected outcome under a null model of no interaction (Fig. 1d,e). Deviations from this expectation were quantified and corrected for floor effects (Fig. 1f and Extended Data Fig. 2d,e), and then aggregated at the gene-pair level to generate a GI score. Gene pairs with greater-than-expected depletion were assigned negative GI scores, and those with less-than-expected depletion yielded positive GI scores (Fig. 1d). To assess confidence, GI scores were modelled as a mixture of typical deviations (normal distribution) and deviations outside the norm (uniform distribution). Gene pairs with interaction probabilities of 0.5 < p ≤ 0.95 were designated likely interactions, and those with p > 0.95 confident interactions.

The Msm GI map revealed nearly 8,500 interactions (p > 0.5), with negative GIs outnumbering positive ones (Fig. 1g–i and Supplementary Table 2). This 2.7% hit rate across gene pairs aligns with previous GI studies3. Most genes interacted with fewer than 15 partners, but a subset exhibited high connectivity, consistent with hub-like behaviour that has been observed in other organisms, such as yeast3 (Fig. 1i). Out of 787 targeted genes, 64 (8.1%) showed no interactions under these conditions. Notably, GI strength did not correlate with gene vulnerability10, confirming that our fitness-based sgRNA selection accounts for gene vulnerability (Extended Data Fig. 2f). sgRNAs targeting the same gene exhibited highly correlated interaction profiles (Fig. 1j), underscoring the reproducibility of measured GIs.

Illustrative examples confirm that the map captures known or mechanistically plausible functional relationships. We observed a strong negative GI between the Hsp40 chaperones dnaJ1 and dnaJ2 (Fig. 1e), consistent with their known synthetic lethal relationship21. Another strong negative GI, between cwlM and murA (Fig. 1k), is consistent with the proposed role of CwlM in binding to MurA, stimulating its activity, and promoting peptidoglycan precursor synthesis22.

GI profiles reveal gene function

With individual GIs providing initial confidence in the dataset, we next sought a global assessment of the functional signal encoded in the map. To this end, we analysed correlations among GI profiles, because profile similarity often reflects similar gene function3.

Pairwise correlation and clustering revealed that genes with related biological functions tended to group together (Fig. 1l). We systematically annotated the tree, identifying 63 functionally coherent clusters containing 2 to 48 genes each. Representative clusters spanned functions including cell-wall synthesis, protein degradation, cell division, cofactor and nucleotide synthesis, the ETC, and protein folding and ribosome rescue (Fig. 1l, top to bottom, left to right). Several observations highlight the biological relevance of these clusters. First, cwlM and murA share not only a strong negative GI (Fig. 1k), but also highly correlated GI profiles22 (Fig. 1l). Second, the two genes most closely associated with the mycolic acid transporter mmpL3—pgfA and ttfA—were previously shown to interact physically with MmpL3 and support mycolic acid transport23,24. Third, the transcription factors whiA and whiB2, which are involved in cell division and act in a co-dependent manner in Streptomyces, clustered with each other and with ftsZ, consistent with their established roles25 as transcriptional activators of ftsZ. Many clusters also contained previously uncharacterized genes, as well as functionally annotated genes with unexpected correlations, suggesting unrecognized biological connections. Barring technical limitations (for example, polar effects), these clusters offer functional predictions by placing genes in the context of their network relationships.

Essential-gene GI map in Mtb

Encouraged by the validity of the Msm map, we next extended our approach to Mtb. Following the same experimental blueprint, we constructed a dual-sgRNA library containing up to two sgRNAs per gene with reproducible, intermediate fitness costs (Supplementary Table 1). The library includes 1,483 sgRNAs targeting 762 genes, most of them essential for in vitro Mtb growth, plus 10 NT controls, yielding about 2.2 million sgRNA permutations. The library was cloned in E. coli, validated by deep sequencing (Extended Data Fig. 2a) and transformed into an H37Rv Mtb strain containing a chromosomally integrated Sth1 dCas9 cassette. Competitive growth and barcode sequencing were performed as in Msm.

As in Msm, growth phenotypes were highly reproducible across biological replicates (Extended Data Fig. 2g,h). Dual-NT, single-targeting and dual-targeting strains depleted on average in the expected order, from least to most (Fig. 2a). sgRNA position within the plasmid had a greater effect than it did in Msm (Fig. 2b), but the effect was modest and deemed acceptable, because our GI model accounts for sgRNA position. GIs were quantified, corrected for floor effects (Fig. 2c and Extended Data Fig. 2i,j) and aggregated at the gene-pair level to generate a GI score and confidence estimate, as in Msm.

Fig. 2: Comprehensive GI map of essential genes in Mtb.

a, Distribution of observed L2FC at generation 22 for dual-NT, single-targeting and dual-targeting constructs, as in Fig. 1b. b, Correlation of observed L2FC at generation 22 for single-targeting constructs differing only in sgRNA position. c, Expected versus observed L2FC at generation 25 for all dual-targeting constructs. Black line, x = y; red line, GAM fit. GI scores were computed as the median GAM residual per gene pair. d, Distribution of GI scores across all gene pairs. Shaded regions denote statistically significant (p > 0.5) interactions. e, Heat map of GI scores for all gene pairs, hierarchically clustered by Pearson correlation of GI profiles. Values beyond ±5 are clipped. f, Number of positive and negative GIs (p > 0.5) per gene. g, Distribution of pairwise GI profile correlations among all sgRNAs, and among sgRNAs targeting the same gene. h, GI scores for all genes paired with dnaJ1 (left) or with dnaJ2 (right). i, GI scores for all genes paired with murA (left) or with cwlM (right). j, GI scores for all genes paired with secA1 (left) or with secF (right). k, Left, dendrogram of all genes, clustered by GI profile correlation. Branches are coloured by nodes enriched (FDR < 0.005) for GO terms by Fisher’s exact test. Right, representative clusters (from top to bottom: fatty acid biosynthesis, arginine biosynthesis, nucleotide biosynthesis, folate metabolism, thiamine metabolism, menaquinone biosynthesis and protein degradation).

The Mtb GI map revealed nearly 6,500 interactions (p > 0.5), with negative GIs outnumbering positive ones (Fig. 2d–f and Supplementary Table 3). This 2.2% hit rate across all gene pairs parallels the Msm map. Most genes interacted with fewer than 12 partners, but a subset exhibited high connectivity (Fig. 2f), and 3.4% (26/762) of targeted genes showed no interactions under these conditions. As in Msm, sgRNAs targeting the same gene exhibited highly correlated interaction profiles (Fig. 2g), and GI strength did not correlate with gene vulnerability (Extended Data Fig. 2k). We observed strong negative GIs between dnaJ1 and dnaJ2 (ref. 21) and between cwlM and murA (ref. 22) (Fig. 2h,i). Further strong negative GIs recapitulated known or mechanistically plausible functional relationships, including (1) transcription factors and their targets (Extended Data Fig. 3a–c); (2) partially redundant enzymes (Extended Data Fig. 3d,e); and (3) functionally related genes within shared biological processes (Fig. 2j and Extended Data Fig. 3f). Reflecting the biological coherence of the data, clustering by GI profile similarity grouped genes with related functions (Fig. 2k).

To test platform reproducibility, we repeated the Mtb GI screen using a slightly higher concentration of anhydrotetracycline (ATc) to ensure saturating induction of dCas9 and sgRNA (Supplementary Table 3). We refer to the original and replicate screens as experiment 1 and experiment 2, respectively. The second screen showed slightly greater positional variability but also more pronounced depletion of single- and dual-targeting constructs (Extended Data Fig. 4a,b), resulting in a wider distribution of GI scores (Extended Data Fig. 4c–e) and a higher signal-to-noise ratio. Key benchmarks were again met, including concordance among replicates (Extended Data Fig. 2l–o), high GI profile similarity among sgRNAs targeting the same gene (Extended Data Fig. 4f) and a lack of correlation between GI score and gene vulnerability (Extended Data Fig. 2p). Most importantly, the replicate map recapitulated the functional information in the original. Summary metrics, including the overall hit rate (2.3%) and per-gene interaction number (Extended Data Fig. 4g), were comparable; specific GIs were reproduced across the two screens (Extended Data Figs. 3g–l and 4h,i); and functionally related genes clustered together (Extended Data Fig. 4j).

GI scores were strongly correlated between the Mtb experiments (R = 0.82; Extended Data Fig. 4k). However, concordance between interaction probabilities was lower (Supplementary Table 3), particularly near the boundary between the null and interaction distributions, where experiment-specific fits diverged. To reconcile these differences, we extended our statistical framework to jointly model GI scores from both experiments (Extended Data Fig. 4k). The model assigns each gene pair a joint probability and, when sufficiently confident, classifies the interaction as negative, positive or discordant. Joint modelling identified confident GIs for approximately 2.5% of Mtb gene pairs (p > 0.95) and likely GIs for 4.4% (p > 0.5).

Given the substantial but incomplete overlap between the two maps, we retained data from both screens. Unless otherwise specified, the individual Mtb GI scores and downstream analyses presented in the following sections are based on experiment 2, and the joint model provides an integrated assessment of interaction confidence across both experiments.

Overview of mycobacterial GIs

Having established platform robustness in both Mtb and Msm, we analysed GIs involving homologous genes targeted in both libraries. Consistent with comparative GI work in budding and fission yeast26, homologues showed a modestly higher GI profile correlation than did non-homologues (Fig. 3a,b), reflecting a mixture of conserved (for example, Figs. 1e,k and 2h,i) and non-conserved (for example, Extended Data Fig. 5a–c) interactions. A global comparison of GI scores for homologous gene pairs revealed limited correlation across species (R = 0.26; Extended Data Fig. 5d), which increased substantially when the analysis was restricted to gene pairs sharing the same biological process annotation (R = 0.57; Extended Data Fig. 5e and Supplementary Table 4). Likewise, within-process GIs were conserved at a higher rate than the overall set of interactions (Extended Data Fig. 5f). Together, these findings indicate that mycobacteria recapitulate the pattern of GI conservation observed in budding and fission yeast, in which interactions that link genes in the same biological process are preferentially conserved relative to interactions that link distinct processes26.

Fig. 3: Overview of mycobacterial GIs and integration with PPI predictions.

a,b, GI profile correlations between Msm and Mtb homologues in experiment 2 (a) and between identical genes across Mtb experiments (Mtb experiment 1 versus Mtb experiment 2; b). Each of the 100 black lines in each panel shows the density obtained using a random sample of non-homologous (a) or non-identical (b) gene pairs. c, Heat map of log2 odds ratio (OR), testing for enrichment or depletion of significant negative GIs (p > 0.5) within pathway pairs in Mtb. Gene pairs were categorized by GI presence or absence and pathway-pair membership, and association was assessed by Fisher’s exact test. Positive and negative log2(OR) values indicate more and fewer GIs than expected, respectively. Only results with |log2(OR)| > 2 and q < 0.0001 are shown. Hollow circles indicate pathway pairs that fail these thresholds after removing possible polar effects (gene pairs ≤ 5 genes apart). See Extended Data Fig. 5g for all significant (q < 0.05) pathway pairs. d,e, Distribution of s.c. ipTMs for co-folded Mtb protein pairs with mean s.c. ipTM ≥ 0.2 (d) and for all protein pairs (e). Black dots and grey lines represent the mean and s.d. of s.c. ipTMs across all pools. f, Distribution of PPI confidence across bins of GI score and GI profile correlation in Mtb. Gene pairs were binned by GI score (left) or by GI correlation (right), and the fraction of corresponding protein pairs falling into each mean s.c. ipTM category is shown. Per-bin gene-pair counts are indicated. PPI confidence: none (mean s.c. ipTM < 0.05), low (0.05 ≤ mean s.c. ipTM < 0.2), medium (0.2 ≤ mean s.c. ipTM < 0.4) and high (mean s.c. ipTM ≥ 0.4). g–n, AlphaFold-Multimer models and GI and PPI metrics for selected Mtb pairs: MurA and CwlM (g), DnaN and DnaE1 (h), DesA2 and DesA1 (i), EtfD and EtfB (j), Rv3212 and Rv1002c (k), AftD and TmaT (l), WecA and Rv3355c (m) and AftB and Rv3035 (n).

We next examined the functional organization of Mtb GIs by testing for enrichment of interactions within and between biological processes (Supplementary Table 4). We focused on negative GIs, because they are more abundant in our dataset and more likely to reflect functional relationships, following the precedent in yeast3. Gene pairs comprising genes in the same biological process were enriched for negative GIs (on-diagonal; Fig. 3c and Extended Data Fig. 5g), consistent with limited tolerance for simultaneous knockdown within essential pathways. Negative GIs were also enriched among gene pairs spanning distinct biological processes—for example, genes involved in synthesizing different components of the cell envelope (such as arabinogalactan and lipoarabinomannan), and genes involved in DNA replication and DNA repair (off-diagonal; Fig. 3c and Extended Data Fig. 5g). Beyond these process-level patterns, gene pairs with highly correlated GI profiles were themselves enriched for negative GIs (Extended Data Fig. 4l), consistent with the expectation that genes acting in the same pathway are more likely to share interaction partners and interact with each other3.

Coupling GIs with protein interactions

To investigate the molecular basis of GIs, we asked how frequently they reflect physical associations. Because no large-scale experimental protein–protein interaction (PPI) map exists for Mtb, we generated PPI predictions using pooled AlphaFold3 co-folding4,27 across the approximately 600,000 gene pairs screened in the Mtb and Msm GI libraries. We identified 15,240 candidate PPIs using a size-corrected interface predicted template modelling score (s.c. ipTM) cut-off ≥ 0.05 (Fig. 3d,e, Extended Data Fig. 6a–h and Supplementary Tables 2 and 3). Because size correction systematically reduces ipTM scores, this threshold is necessarily permissive, yet the overall hit rate (2.6%) remains low. Many of these candidates achieve higher ipTM scores when evaluated pairwise with AlphaFold-Multimer28. Consistent with observations in yeast3, gene pairs with negative GIs or highly correlated GI profiles were enriched for predicted physical interactions (Fig. 3f). Nevertheless, many lacked a predicted PPI, because GIs can arise from pathway-, process- or network-level relationships without direct binding. Overall, this resource identifies gene pairs that are both genetically and physically linked, providing a framework for predicting their molecular mechanisms. We present several representative examples spanning both known22,29,30,31 (Fig. 3g–j) and previously unrecognized (Fig. 3k–n) biological connections.

GIs predict novel drug interactions

Previous studies have shown that GIs can predict drug–drug interactions5. To test whether this principle holds in our dataset, we analysed GIs among established drug targets.

One of the strongest negative GIs in both Mtb and Msm involved aspC (alanine aminotransferase16,32) and ddl (d-alanine-d-alanine ligase; Fig. 4a). Although previously unknown, this interaction is mechanistically plausible: AspC is crucial for biosynthesis of l-Ala, which is required by Alr (alanine racemase) to produce d-Ala, the substrate for Ddl (Fig. 4b). aspC also showed a negative GI with alr. Alr and Ddl are the targets of d-cycloserine33, a second-line agent for multidrug-resistant TB. Consistent with the GI data, CRISPRi-based chemical–genetic interaction (CGI) profiling of d-cycloserine identified aspC as one of the top sensitizing hits (Fig. 4c,d and Supplementary Table 5). Moreover, a paper published while we revised this one reported synergy between an AspC inhibitor and d-cycloserine34.

Fig. 4: GIs can predict novel drug interactions.

a, GI scores for all genes paired with aspC in Mtb. b, Pathway schematic illustrating the roles of AspC, Alr and Ddl in peptidoglycan biosynthesis. c,d, Volcano plots from Mtb CGI screens with d-cycloserine (an inhibitor of Ddl and Alr) at concentrations of 6.4 μM (c) and 19.1 μM (d). FDR was calculated using α-Robust Rank Aggregation (RRA) with one-sided tests for enrichment and depletion. e, GI scores for all genes paired with pptT in Mtb. f, Schematic highlighting PptT’s connection to the FAS-II mycolic acid biosynthesis cycle. PptT and the negative interactors discussed in the text are shown in bold. g,i,j,l, Checkerboard assay heat maps showing Bliss scores for the indicated drug pairs in Mtb. Dashed boxes outline the most synergistic area (MSA) and most antagonistic area (MAA). MSA Bliss scores >15 and MAA Bliss scores <–15 indicate synergy and antagonism, respectively. g, The PptT inhibitor 8918 and the KasA and KasB inhibitor thiolactomycin (TLM); i, 8918 and the HadAB and HadBC inhibitor thiacetazone (TAZ); j, 8918 and the HadAB and HadBC inhibitor isoxyl (ISO); l, the EfpA inhibitor BRD-8000.3 and the Pks13 inhibitor mCMX410. n = 3 biological replicates. h, MSA Bliss scores from TLM + 8918 checkerboard assays in reference clinical strains (identified by ITM number): 2018-00083 (83), 2018-00084 (84), 2018-00085 (85), 2018-00088 (88), 2018-00092 (92) and 2018-00094 (94). Bars are coloured by phylogenetic lineage. Dashed line indicates the threshold for synergy. n = 3 biological replicates. k, GI scores for all genes paired with pks13 in Mtb. m, GI scores for all genes paired with pknA (top) or with pknB (bottom) in Mtb. n, Killing assays in Mtb. Left: Vertex compound 49 (V49), a PknA and PknB inhibitor, combined with the ATP synthase inhibitor bedaquiline (BDQ); right: V49 combined with the EmbA and EmbB inhibitor ethambutol (EMB). Colony-forming units (CFU) per ml were quantified after six days (n = 3 biological replicates). Diff. denotes the fold difference in killing between the indicated drug combination and the corresponding highest single agent (HSA) effect; a >100-fold difference defines synergy. Asterisks denote statistical significance (P < 0.001 by Welch’s two-sided two-sample t-test).

Another set of novel, mechanistically plausible interactions involved PptT, a phosphopantetheinyl transferase that activates acyl carrier proteins (ACPs) required for lipid biosynthesis and is an active drug development target35. PptT showed negative GIs with kasA, hadB and hadC (Fig. 4e), which catalyse steps in the FAS-II pathway responsible for mycolic acid biosynthesis (Fig. 4f), and are likewise being pursued as drug targets36,37. Consistent with the GI data, CGI profiling of thiolactomycin, a KasA and KasB inhibitor, identified pptT as a top sensitizing hit (Extended Data Fig. 7a and Supplementary Table 5). Moreover, co-treating Mtb with thiolactomycin and the PptT inhibitor 8918 produced strong synergy in both the H37Rv reference and clinical Mtb strains38 representing diverse lineages (Fig. 4g,h). This synergy was specific: neither thiolactomycin nor 8918 synergized with control drug treatments (Extended Data Fig. 7b–g). Although KasA inhibitors and rifampicin are synergistic39, kasA and rifampicin’s target rpoB did not show a negative GI (Supplementary Table 3). In this case, KasA inhibition is likely to potentiate rifampicin by increasing envelope permeability and intracellular rifampicin exposure, rather than through a target-based mechanism39. Supporting this interpretation, knockdown of rpoB did not sensitize to thiolactomycin in our CGI screen (Extended Data Fig. 7a). These examples show how KasA inhibition can drive synergy through both target-specific and chemotype-specific mechanisms, and how GI mapping can disentangle these modes of action. Combined inhibition of PptT and the HadAB and HadBC dehydratases likewise produced strong synergy (Fig. 4i,j), providing further evidence that negative GIs can predict drug–drug interactions.

To assess whether positive GIs likewise have predictive value, we focused on the pks13–efpA pair. Pks13 and EfpA are essential proteins involved in mycolic acid biosynthesis and lipid translocation, respectively, and both are active targets for TB drug development40,41. The pair showed a strong positive GI (Fig. 4k), predicting drug–drug antagonism, which was confirmed by checkerboard assays of preclinical Pks13 and EfpA inhibitors (Fig. 4l and Extended Data Fig. 7h,i).

As a final proof-of-concept, we examined PknA and PknB, essential eukaryotic-like serine/threonine kinases that regulate Mtb growth and division and are promising drug targets42. pknA and pknB exhibited strong negative GIs with embA and embB (Fig. 4m), which encode arabinogalactan biosynthetic enzymes targeted by the first-line TB drug ethambutol43. Co-treatment with ethambutol and the PknA and PknB inhibitor V49 resulted in borderline synergy in checkerboard assays (Extended Data Fig. 7j), but clear synergy in bacterial killing assays (Fig. 4n). This bactericidal interaction seemed to be specific: pknA and pknB did not exhibit negative GIs with ATP synthase genes (Fig. 4m), and V49 did not enhance killing by the ATP synthase inhibitor bedaquiline44 (Fig. 4n). Although V49 and bedaquiline showed borderline synergy in checkerboard assays (Extended Data Fig. 7k), this might reflect the same uptake-mediated effect proposed for KasA inhibition and rifampicin, rather than a target-based interaction39. Together, these results illustrate how GI mapping can reveal combination effects that are missed by conventional growth-inhibition assays.

Beyond the specific interactions discussed above, the Mtb GI map offers broader opportunities for drug discovery. Strong GIs between currently pursued TB drug targets are rare, consistent with the general scarcity of target-based synergy and antagonism45. Nonetheless, the map reveals thousands of previously unexplored interactions among established and potential drug targets across diverse pathways. We cover several of these interactions in the Supplementary Discussion and Extended Data Figs. 8 and 9. The map also identifies a set of ‘pleiotropic sensitizers’—including ppk, macE, dnaA, rnc, treS, sodA and pknG—that showed strong negative GIs across numerous targets, consistent with findings from our previous CGI studies39. Inhibitors of these pleiotropic sensitizers could potentiate multiple drug classes by globally altering bacterial metabolism, stress responses or regulatory networks, making these genes promising candidates for combination therapy aimed at amplifying the activity of diverse antibacterial agents.

New players in electron transport

Oxidative phosphorylation—the combined activities of the ETC and ATP synthase—has emerged as an important antibiotic target in Mtb and related pathogens. Clinically validated examples include the US Food and Drug Administration (FDA)-approved ATP synthase inhibitor bedaquiline44 and telacebec (Q203)46, a clinical drug candidate that targets complex III. In mycobacteria, complex III2IV2 (Fig. 5a) is the primary terminal oxidase that establishes the proton gradient driving ATP synthesis47. This supercomplex (larger than 700 kDa) consists of 24 protein subunits47 that position 18 redox cofactors to enable electron transfer coupled with proton translocation. In eukaryotes, numerous proteins chaperoning the maturation of complex III and IV have been identified, and defects in these assembly factors cause severe hereditary diseases48. In mycobacteria, the mechanisms of complex III2IV2 assembly and cofactor incorporation remain mostly unknown49.

Fig. 5: Discovery of genes that are important for the function of the mycobacterial ETC.

a, Mycobacterial complex III2IV2 (Protein Data Bank (PDB) ID: 9DM1). Purple, complex III; green, complex IV; blue, malate quinone oxidase; pink, superoxide dismutase; black, bound cofactors. MQ, menaquinone. b,c, GI profile correlations for complex III2IV2 genes and strongly correlated genes of unknown function in Msm (b) and Mtb (c). Homologous genes are aligned. d,e, GI scores for genes paired with cytochrome bd oxidase genes cydA and cydB in Msm (d) and Mtb (e). f–i, Spotting assays of CRISPRi strains in wild-type and cytochrome bd knockout (Δbd) backgrounds in Msm (f,h) and Mtb (g,i). mtrA serves as a negative control. Representative image shown from one of two (Msm) or three (Mtb) biological replicates. j,k, Heat maps of s.c. ipTM scores in Msm (j) and Mtb (k). l, Methylene blue oxygen consumption assays of Δbd Mtb strains with knockdown of the indicated genes or treatment with 400 nM Q203 (a complex III2IV2 inhibitor). ATc was added five days before the assay. Clear, normal oxygen consumption; blue, respiratory defects. Imaged three days after adding methylene blue and sealing the vials. Representative image shown from one of three independent experiments. m, Minimum inhibitory concentrations yielding 50% growth inhibition (MIC50) relative to each strain’s no-drug control for Q203 in Δbd Mtb CRISPRi strains. ATc was added five days before the assay. Bars show point estimates from a four-parameter log-logistic dose–response model; error bars show 95% Wald (delta-method) confidence intervals. n = 3 biological replicates. n, AP–MS results for tagged and untagged Ms2764. SAINTexpress analysis was used to compare two independent purifications with two untagged control purifications to calculate an L2FC and FDR for each protein detected. n = 2 biological replicates. o, IMV acidification assays for the indicated strains. Fluorescence is quenched when protons are pumped into the IMVs after the addition of an electron donor and recovers after addition of the K+:H+ antiporter nigericin. n = 3 biological replicates. RFU, relative fluorescence units; WT, wild type.

We identified a cluster of genes of unknown function that strongly correlated with genes encoding subunits of complex III2IV2 and with genes involved in cofactor biogenesis, in both Msm and Mtb (Figs. 1l and 5b,c and Extended Data Fig. 10a). Many of these uncharacterized genes also showed negative GIs with the alternative, partially redundant terminal oxidase, cytochrome bd oxidase (cyt. bd), encoded by cydA and cydB (Fig. 5d,e). Synthetic lethality with cyt. bd is a hallmark of defects in complex III2IV2 in mycobacteria50. We thus hypothesized that these genes contribute to the assembly or function of complex III2IV2. Several lines of evidence support this hypothesis: (1) independent single-strain knockdowns using sgRNAs not included in the screen validated the GI results in both Msm and Mtb (Fig. 5f–i); (2) complementation experiments in Msm ruled out off-target effects and established the functional equivalence of Mtb homologues (Extended Data Fig. 10b); (3) these genes showed negative GIs with each other in Mtb (Extended Data Fig. 10c); (4) AlphaFold3 predictions suggest physical interactions among several of these genes and with CtaD, a complex IV subunit (Fig. 5j,k and Extended Data Fig. 10d); and (5) knockdown of these genes in strains lacking cyt. bd reduced oxygen consumption, an indicator of complex III2IV2 activity (Fig. 5l and Extended Data Fig. 10e–g), and increased sensitivity to Q203 to varying extents (Fig. 5m and Extended Data Fig. 10h).

To extend these findings biochemically, we focused on ms2764, selected for its AlphaFold3-predicted interactions and the tractability of generating a functional 3×Flag-tagged construct with correct membrane topology. To our knowledge, there are no previous studies of ms2764 or its Mtb homologue rv2698, which encode membrane proteins containing a DUF3093 domain that seems to be restricted to Actinobacteria. Affinity purification of Ms2764–3×Flag followed by mass spectrometry (AP–MS) showed co-purification of complex III2IV2 subunits and several other uncharacterized genes in the cluster (Fig. 5n). Motivated by this evidence of physical interaction, we generated an ms2764 knockout and assayed respiratory activity in inverted membrane vesicles (IMVs). As shown previously51, NADH-driven IMV acidification can occur even when complex III2IV2 is inhibited or absent, whereas succinate-driven IMV acidification requires functional complex III2IV2. Similar to IMVs from a ∆qcrB strain, ∆ms2764 IMVs were capable of NADH-driven acidification but not succinate-driven acidification, indicating loss of complex III2IV2 activity (Fig. 5o) and providing biochemical evidence that Ms2764 contributes to complex function.

Closer inspection of these uncharacterized genes revealed that one (rv2235 or ms4311; here renamed surf1) contains a well-conserved SURF1 domain, known in other species to facilitate the incorporation of haem a3 into CtaD (ref. 49). A second gene (rv0102 or ms4702; here renamed ctiP) probably encodes a distant CtiP homologue (25% sequence identity) that, in Corynebacterium, seems to mediate CuB insertion into CtaD (ref. 49). Consistent with this proposed role, ctiP is encoded adjacent to chalkophore biosynthesis genes in Mtb, which supply copper to complex IV under copper-limiting conditions and during infection52. This genomic context suggests that CtiP collaborates with chalkophores to metallate complex IV. rv1342c (ms4900) and lprD are induced under iron starvation53, consistent with the iron dependence of complex III2IV2 activity, whereas rv1457c (ms3118) and rv1458c (ms3119) are located immediately upstream of ctaA, which catalyses the final step of haem A biosynthesis. For the remaining genes, no precedent links them to the ETC. They encode predicted integral membrane proteins (rv2698 (ms2764) and rv1083 (ms5260)) and lipoproteins (lppL, rv2980 (ms2396)). These uncharacterized genes might therefore participate in cofactor trafficking or installation or represent previously unrecognized subunits of complex III2IV2. Although further studies are required to define their molecular roles, these findings show how our GI platform enables hypothesis generation for previously uncharacterized genes.

Discussion

A large portion of the Mtb genome remains poorly understood. Here, to help close this gap, we developed a scalable approach for mapping genetic interactions in mycobacteria, focusing on essential genes because they underpin core cellular processes and represent the targets of most antibiotics. Across the model species Msm and the pathogen Mtb, we screened more than four million sgRNA permutations and identified more than 15,000 GIs. This constitutes the first comprehensive GI map of essential genes in Mtb—and, to our knowledge, the first such map for essential genes in any organism beyond budding yeast3. Together with other large-scale resources41,54,55,56,57—including transcriptomic and metabolomic compendia, systematic gene knockdown and morphological profiling, curated TnSeq datasets and chemical–genetic profiling—these datasets unlock new opportunities for dissecting mycobacterial biology and guiding the rational design of drug combinations.

Three applications of our maps merit particular attention. First, they can identify interactions between drug targets, enabling prediction of synergistic or antagonistic drug combinations. Second, GI profile similarity can reveal functionally related genes. Third, GI mapping can uncover surprising links between well-characterized genes and distinct biological processes. Although some unexpected associations might reflect technical artefacts (for example, off-target or polar effects), many are likely to reveal genuine biology. For example, the ribonuclease III gene rnc, best known for rRNA maturation58 and CRISPR RNA processing59, clusters with cell-division genes in both Mtb and Msm (Fig. 1l). This association aligns with observations60 in Corynebacterium glutamicum, in which Rnc cleaves the mRNA of the cell-division regulator MraZ.

Shortening the duration of TB treatment remains a central objective in TB drug development. One promising strategy is to exploit interactions between drug targets. GI mapping, already a powerful approach in cancer therapeutics3, offers analogous opportunities for TB. First, GI maps can highlight synergistic and antagonistic target combinations to exploit and to avoid, guiding the rational selection of drug combinations. We validated four synergistic compound pairs (8918 + TLM, 8918 + TAZ, 8918 + ISO and V49 + EMB) and one antagonistic pair (BRD-8000.3 + mCMX410) predicted by our GI maps. We also identified and validated numerous negative GIs that link established drug targets to genes not currently targeted pharmacologically. These findings nominate the corresponding gene products as potential targets for future drug development. GI data can also illuminate the mechanistic basis of drug–drug synergy—distinguishing, for example, uptake-based and target-specific interactions. Moreover, because pairwise interactions might predict higher-order interactions61, these datasets provide a foundation for exploring regimen-level effects. This information represents a key piece of the puzzle, although other factors beyond drug synergies contribute to therapeutic success62. Second, GI maps can help to ‘de-orphan’ potential drug targets by predicting their function—an essential step, because understanding a target’s biochemical role is a prerequisite for target-based drug discovery. Third, GI profiles provide genetic signatures for inferring the mechanisms of action of uncharacterized compounds. A close resemblance between a compound’s CGI profile and a gene’s GI profile could implicate that gene as the compound’s primary target5. Two caveats are worth noting. First, CRISPRi-mediated target depletion is not equivalent to small-molecule inhibition. CRISPRi approximates noncompetitive inhibition; it does not capture the full spectrum of pharmacological mechanisms. CRISPRi perhaps most closely resembles targeted protein degradation, as exemplified by emerging BacPROTACs63. Consequently, GIs probably exhibit only partial overlap with CGIs. Second, drug interactions are concentration dependent, and the fitness effects observed here might not correspond to the concentration window in which a drug interaction occurs.

Elucidating gene function is fundamental to understanding Mtb biology. Although individual pairwise GIs can be informative, they often do not reveal the broader functional relationships between genes. GI profile correlations address this limitation by integrating GIs across gene pairs, enabling the identification of functionally related genes through shared interaction patterns. Indeed, clustering genes by profile similarity in our maps recovered coherent pathways, and placing a gene within a pathway is often a major step towards defining its role. However, pathway assignment alone rarely yields specific molecular hypotheses. To help bridge this gap and clarify the relationship between GIs and PPIs, we generated PPI predictions for all essential genes in Mtb and Msm. We found that genes exhibiting negative GIs, as well as those with correlated GI profiles, are enriched for predicted physical interactions, increasing confidence that these genetic connections reflect real biological relationships. However, the overlap between GIs and PPIs remains incomplete, underscoring that each dataset captures distinct but complementary dimensions of gene function. To illustrate how GI data can help to assign functions to uncharacterized genes, we examined a subset of genes of unknown function whose interaction profiles strongly correlated with those of ETC components. Through orthogonal assays, we confirmed their association with complex III2IV2, providing a first step for further investigation.

Our GI maps were generated using standard laboratory strains and culture conditions. Even in this form, our maps are relevant to clinical Mtb strains, enabling the prediction of strain-specific vulnerabilities. For instance, we previously showed39 that sublineage L1.2.1, which is endemic to the Philippines, Thailand and Indonesia, carries a loss-of-function mutation in whiB7. The GIs observed here suggest that this background is particularly sensitive to inhibition of histidine and alanine biosynthesis16. Beyond these predictions, our maps provide a foundation for GI mapping across diverse genetic backgrounds. Such studies could help to explain why, despite relative genetic homogeneity, Mtb strains exhibit differences in clinically relevant traits, including transmission and response to antibacterial agents64. At the same time, how growth environment reshapes GIs in Mtb remains an open question. Studies in budding yeast indicate that although most GIs are preserved across growth conditions, approximately 5–24% of them are condition specific65. Thus, defining how genetic and environmental contexts rewire Mtb’s GI network will be crucial to understanding its adaptive physiology and identifying condition-dependent drug vulnerabilities.

This work builds on pioneering efforts to map GIs in E. coli66 and budding yeast2,3 and expands on earlier efforts in Mtb9. By generating comprehensive GI maps of essential genes in Mtb and Msm, we provide a rich resource with broad utility for elucidating Mtb biology and guiding the rational design of combination therapies. All data are publicly available through an interactive web server (https://pebble.rockefeller.edu/pages/gi/). These maps lay the groundwork for probing how GIs change across environments—such as under antibiotic pressure or in infection-mimicking conditions. Moreover, this framework can be extended to encompass non-essential genes, ultimately enabling genome-wide, all-by-all GI maps. The experimental and computational strategies described here also provide a blueprint for systematic GI mapping in other pathogens, with the potential to accelerate therapeutic discovery and combat the growing threat of drug-resistant infectious diseases.

Limitations of the study

There are six main limitations to these GI maps. First, this dataset represents a first-generation GI map in mycobacteria and, as such, inevitably includes both false positives and false negatives. Future iterations incorporating improved library design, expanded replication and refined statistical modelling should increase confidence in individual interaction calls. Second, by design, this map focuses on essential genes, reflecting current experimental limits: extending the present approach to an all-by-all Mtb GI map would require testing around 64 million sgRNA permutations. Such comprehensive maps remain a goal for future work. Third, CRISPRi-mediated polar effects can produce spurious GIs and correlations. This issue is partially mitigated by the fact that genes within operons often participate in related pathways, although it remains an inherent limitation. Fourth, a small subset of genes (about 25 of 787 in Msm) exhibited numerous strong correlations that are unlikely to reflect genuine biological relationships. The cause of these correlations is unclear, and caution is warranted when investigating these genes. Fifth, although we observe overlap in ‘hub genes’ between our GI data and published CGI datasets39, some of these might represent technical artefacts. Finally, dCas9 can act as a physical barrier to DNA replication67, potentially producing artefactual negative GIs with DNA-replication-associated proteins.

Methods

Bacterial strains

Mtb strains (unless specified otherwise) are derivatives of H37Rv. Clinical strains were acquired from the BCCM/ITM reference set of clinical strains38 and are referred to by their ITM number. Msm strains are derivatives of mc2155. E. coli strains are derivatives of DH5α (NEB, C2987).

Mycobacterial cultures

Mycobacteria were grown at 37 °C in Difco Middlebrook 7H9 broth (BD, 271310) or on 7H10 agar (BD, 262710) supplemented with 0.2% (7H9) or 0.5% (7H10) glycerol, 0.05% Tween-80 and either 1× albumin dextrose catalase (ADC) (for Msm) or 1× oleic acid albumin dextrose catalase (OADC) (for Mtb). Antibiotics (10–20 μg ml−1 kanamycin, 5–20 μg ml−1 zeocin, 50 μg ml−1 hygromycin and 25 μg ml−1 nourseothricin) were added as needed to construct and maintain strains and libraries, and 100–500 ng ml−1 ATc was used to induce CRISPRi. Msm cultures were grown shaking in sealed vessels. Mtb cultures were grown standing in tissue culture flasks with 5% CO2.

Strain construction

Knockout strains

Mtb and Msm deletion mutants lacking cytochrome bd oxidase were derived from the corresponding wild-type strains. The Mtb deletion mutant, in which the cydABDC operon is replaced with a hygromycin resistance cassette, has been described previously68. To construct an analogous Msm mutant, ORBIT69 was used to delete both the cydAB and appCB operons, which encode confirmed and putative bd-type oxidases, respectively. The resulting strain carries hygromycin and zeocin resistance cassettes, with PCR confirming successful replacement of the endogenous loci.

The Msm strains ∆qcrB and ∆ms2764 were prepared using the ORBIT method69. To generate the ∆qcrB strain, wild-type Msm containing plasmid pKM444 (Addgene, 108319) was transformed with the payload plasmid pKM464 (Addgene, 108322) and the oligonucleotide TCGTCGAACCCGTCGGACCGGCATTCTGGGAGCGCAAATCATGAGTCCTGACTTTGCGAAACTCGCCGCCGGTTTGTCTGGTCAACCACCGCGGTCTCAGTGGTGTACGGTACAAACCGGCAACGGGGAGACCAACGGTCATCACTGATCGCTGAGATACTCGGATCGCCGCAATTCCTCTTCGGAGG. To generate the ∆ms2764 strain, wild-type Msm containing plasmid pKM444 was transformed with a modified pKM464, in which the attP/attB core sequence is replaced with GCGTGCTC, and the oligonucleotide CAGAGAGGTGGACACGACGTCACAGATTACTCTGAGCCGCGTGTCGGACACGCGCGCAACCACCCAAACCGGTTTGTCTGGTCAACCACCGCGTGCTCAGTGGTGTACGGTACAAACCGATCGCCTGCTGGCGGCGTTGAACGCCTGATCGAATCGGTCTGCGGCCGGTCAGGCCGCGCAGTCGGTGC. Transformants of ∆qcrB and ∆ms2764 were screened for payload integration by selection with 50 µg ml−1 hygromycin. Successful deletion of qcrB was confirmed by colony PCR using primers GTTCGACGTTTGGAATTCG and CCTGGTATCTTTATAGTCCTGTCG, and successful deletion of ms2764 was confirmed by colony PCR using primers GAAACCTTGATCTCACCGCG and GACGAATTCACCTGCTCCAG.

Epitope-tagged strains

An Msm strain encoding a 3×Flag tag at the C terminus of Ms2764 was prepared using the ORBIT method69. In brief, wild-type Msm containing plasmid pKM444 was transformed with the payload plasmid pSAB4170 and the oligonucleotide GGACGACCCCACCCCGTACTGGCTGGTCAGCACTCGTCACCCGGATCGCCTGCTGGCGGCGTTGAACGCCGGTTTGTCTGGTCAACCACCGCGGTCTCAGTGGTGTACGGTACAAACCTGATCGAATCGGTCTGCGGCCGGTCAGGCCGCGCAGTCGGTGCAGATCATCACGCCGTTCTTCTCGCTGGCGA. Transformants were screened for integration of pSAB41 by selection with 50 µg ml−1 hygromycin followed by colony PCR using primers GAGGTGCCGAAATCGGCGAAGT and CCTGGTATCTTTATAGTCCTGTCG.

Plasmid construction

All plasmids are listed in Supplementary Table 6.

CRISPRi constructs

CRISPRi plasmids were cloned as described previously71 using plRL61 for Msm and plRL58 for Mtb10. These backbones support L5-integration of the complete CRISPRi system and differ by only two nucleotides in the dCas9 Shine–Dalgarno sequence. The substitutions are predicted to reduce dCas9 translation efficiency in Mtb relative to Msm, reflecting the lower expression level required for effective CRISPRi in Mtb. To make each CRISPRi construct, the appropriate backbone was digested with BsmBI (NEB, R0739) and gel-purified before ligation (T4 DNA ligase; NEB, M0202) with annealed oligonucleotides containing the sgRNA targeting sequence (Supplementary Table 6) and compatible overhangs. Ligation products were transformed into chemically competent E. coli cells (NEB, C2987) and selected on LB plates containing kanamycin (50 µg ml−1). Plasmids were isolated from single colonies, and the sgRNA region was verified by Sanger sequencing. Validated plasmids were then co-transformed with plRL19 (L5 integrase helper vector) into mycobacteria as described previously10,71, and transformants were selected with kanamycin (20 µg ml−1).

CRISPRi-resistant complementation constructs

CRISPRi-resistant complementation constructs were cloned as described previously71 into the Tweety-integrating backbone plRL9116. Each construct was assembled from PCR-amplified fragments (Q5 High-Fidelity 2× Master Mix; NEB, M0492) using NEBuilder HiFi DNA Assembly Master Mix (NEB, E2621). Fragments were amplified from plasmid or genomic DNA using custom oligonucleotides designed to introduce (1) 20 bp of overlap between adjacent fragments and (2) silent mutations in the protospacer adjacent motif (PAM) site and seed sequence of the target gene to prevent sgRNA recognition. Gel-purified fragments were incorporated into DraIII-digested and gel-purified plRL91 using a 2:1 or 1:1 fragment-to-backbone ratio, with a 60-min incubation at the recommended temperature. All complementation alleles were expressed from the hsp60 promoter. Assembly products were transformed into chemically competent E. coli cells and selected on LB plates containing carbenicillin (100 µg ml−1). Plasmids were isolated from single colonies and verified by whole-plasmid sequencing (Plasmidsaurus). Validated plasmids were then co-transformed with plRL62 (Tweety integrase helper vector) into mycobacteria as described previously16,71, and transformants were selected with nourseothricin (25 μg ml−1).

Split-plasmid system for CRISPRi libraries

For cloning GI libraries, a split-plasmid CRISPRi system was developed, in which the invariant and variable components of the CRISPRi machinery are carried on separate plasmids that integrate at distinct genomic sites. Sequential transformation of mycobacteria—first with the invariant-element plasmid carrying dCas9, then with the variable-element plasmid carrying the sgRNA library—provides two advantages during library transformation: (1) the reduced size of the library plasmid improves transformation efficiency; and (2) transformation into a clonal dCas9-expressing population minimizes the emergence of CRISPRi-resistant mutants10.

The specific invariant-element plasmids used in this study were plRL252 (Msm) and plRL195 (Mtb). Each carries (1) dCas9 under the control of an optimized, synthetic Tet repressor (TetR)-regulated promoter; (2) a mycobacterial codon-optimized Tet repressor; (3) a Giles attachment site; and (4) a zeocin resistance cassette. Construction of these plasmids was accomplished using NEBuilder HiFi DNA Assembly (described above), with amplification of the dCas9–TetR fragment from the previously described plasmids plRL117 (Msm) or plRL2 (Mtb)10 and assembly into MluI-digested (NEB, R3198) and ClaI-digested (NEB, R0197) plRL12516. Like plRL61 and plRL58 (described above), plRL252 and plRL195 differ from each other by only two nucleotides in the dCas9 Shine–Dalgarno motif, which was engineered for optimal dCas9 expression in each species. Plasmids plRL252 and plRL195 were co-transformed with plRL40 (Giles integrase helper vector) into wild-type Msm and Mtb, respectively, as described previously16,71, and transformants were selected with zeocin (20 µg ml−1).

Complementing the invariant-element plasmids, the variable-element plasmid plRL389 served as the backbone for cloning both the Msm and the Mtb GI libraries. It contains (1) a synthetic TetR-regulated promoter driving sgRNA expression; (2) an L5 integrase and attachment site; and (3) a kanamycin resistance (KanR) cassette. plRL389 was derived from plRL210 through multiple cloning steps designed to accomplish the following: (1) removal of the dCas9 and TetR elements, which were relocated to the invariant-element plasmids (see above); (2) synonymous mutation of a XhoI restriction site in the KanR cassette for compatibility with downstream cloning; (3) replacement of the E. coli origin of replication with a predicted higher-yield version; and (4) introduction of restriction and primer-binding sites necessary for downstream cloning and sequencing. These latter elements were inserted directly after the sgRNA scaffold and include, in order: (1) an MfeI restriction site; (2) a landing site for reverse library amplification primers; (3) a XhoI restriction site; (4) an 18-bp barcode placeholder; (5) a SalI restriction site; (6) a landing site for the custom sequencing primer; and (7) a landing site for the common forward library amplification primer. A final restriction site (EcoRI) required for cloning was already present in the parent plasmid directly upstream of the sgRNA cassette. plRL389 was provided to the biotechnology company Genewiz to initiate GI library construction (see ‘CRISPRi library production’).

Note that the standard (unsplit) CRISPRi system was used for the CGI library RLC25, whose sub-libraries RLC3 and RLC14 were cloned into plRL2 and plRL58, respectively (see ‘CRISPRi library production’).

Off-target analysis

With practical constraints limiting the design of GI libraries to two guides per gene, the identification of guides that are least likely to cause off-target knockdown took on new importance. A custom script was applied to each candidate guide to identify off-target matches, defined as instances of perfect complementarity between the 3′ (PAM-proximal) end of the sgRNA sequence and the genome, spanning ten or more base pairs and positioned next to a PAM site other than the intended target. Guides were then stratified into four tiers according to the following criteria: tier 1, no off-target matches; tier 2, no off-target matches targeting the non-template strand of a gene; tier 3, no off-target matches targeting the non-template strand of a gene with a PAM sequence conferring fold repression ≥ 42.2 in a luciferase knockdown assay13; and tier 4, all remaining guides.

Moving from the first tier to the fourth, the risk of off-target activity increases. The focus on non-template-targeting matches reflects previous data establishing the greater efficacy of non-template- versus template-targeting CRISPRi11,13. The PAM strength cut-off of 42.2 was chosen because weaker PAM sequences rarely produce detectable off-target effects, as inferred from depletion of guides targeting non-essential genes with off-target matches in essential genes.

CRISPRi library design

Msm GI library

The Msm GI library was designed to target all essential genes, in addition to Msm homologues of Mtb in vitro essential genes. Essentiality was defined as the union of TnSeq calls72 and CRISPRi calls10; in the latter case, essentiality was determined using a mixture model of the estimated maximum depletion rate, βmax, for each gene. This produced a list of 834 genes for which guide selection was attempted.

An automated guide selection pipeline was used to select up to two sgRNAs per gene. For genes essential by CRISPRi, sgRNAs were selected in several rounds, with each round becoming more permissive to increase the chance of identifying suitable sgRNAs. The first round of selection was restricted to sgRNAs imposing a specific fitness cost previously determined in a pilot screen to be optimal for detecting GIs: predicted log2-transformed fold change (L2FC) at 25 generations between –5 and –1 in our reference experiment10. In addition, only sgRNAs without off-target matches (off-target tier 1) were considered. In the second round, sgRNAs were permitted to have intergenic and template-targeting genic off-targets (off-target tier 2). In the third round, the guide pool was further expanded to include sgRNAs within a larger L2FC window (–6 to –0.5) and sgRNAs possessing weak non-template-targeting off-targets (off-target tier 3). The final round of selection placed no restrictions on off-target matches (off-target tier 4). In all four rounds, priority was given to sgRNAs closest to the logistic curve describing a gene’s expression–fitness relationship10.

For genes categorized as non-essential by CRISPRi, sgRNAs with the highest predicted strengths10 were selected. To minimize the potential for off-target effects without excessively constraining PAM site availability, a first round of selection was limited to sgRNAs in off-target tiers 1–3. In a second round of selection, the guide pool was expanded to include sgRNAs in off-target tier 4.

For all genes, the selection of length isomers (sgRNAs targeting the same PAM site) was prohibited. In addition, guides that were computationally flagged as outliers and guides introducing restriction sites necessary for library construction were excluded.

Automated guide selections were subjected to minimal manual curation, which involved (1) the replacement of sgRNAs exhibiting unexpected depletion and (2) the exclusion of genes for which only a single guide in off-target tier 4 could be identified.

The final selections consisted of 1,524 sgRNAs targeting 787 genes. The addition of 10 non-targeting control sgRNAs brought the final guide count to 1,534.

Mtb GI library

The Mtb GI library was designed to target all in vitro essential genes, in addition to 72 non-essential genes that were predicted to interact with other genes in the library, are important for infection or are under positive selection in clinical strains. Essentiality was defined as the union of TnSeq calls73 and CRISPRi calls10. This produced a list of 840 genes for which guide selection was attempted.

An automated guide selection pipeline was used to select up to two sgRNAs per gene, similar to the selection process described for the Msm GI library. For genes categorized as essential by CRISPRi, sgRNAs were selected to have a predicted L2FC at 25 generations between –6 and –0.5 (ref. 10). Priority was given to sgRNAs closest to the logistic curve describing a gene’s expression–fitness relationship. For genes that were non-essential by CRISPRi, sgRNAs with the highest predicted strengths were selected. For all genes, guides computationally flagged as outliers and guides introducing restriction sites necessary for library construction were excluded. Automated guide selections were subjected to minimal manual curation, which involved the replacement of sgRNAs exhibiting unexpected depletion.

The final selections consisted of 1,483 sgRNAs targeting 762 genes. The addition of 10 non-targeting control sgRNAs brought the final guide count to 1,493.

Mtb library for CGI profiling

To profile CGIs, an optimized library was designed to target all Mtb genes. This library, designated RLC25, is a combination of two libraries: (1) the previously described library RLC3 (ref. 10), consisting of strong sgRNAs targeting in vitro non-essential genes; and (2) the recently designed library RLC14 (ref. 74), comprising hypomorphic sgRNAs that target in vitro essential genes (union of TnSeq and CRISPRi essentiality calls).

To select sgRNAs for RLC14, behaviour in the reference experiment was used as a starting point, with emphasis on the parameter βe, which models the rate of depletion of a given sgRNA over time10. For each in vitro essential gene, up to 20 sgRNAs were selected. The selection algorithm prioritized depletion rates around βe = −0.14 while also including some stronger and weaker guides. To achieve this, bins representing the 20th percentiles of a normal distribution centred at –0.14 were defined, and one sgRNA was selected from each bin. This approach ensured that most sgRNAs were selected near –0.14 (where the bulk of the distribution lies) while still incorporating guides from the distribution tails. If a bin contained no available sgRNAs, other bins were resampled until a maximum of 20 sgRNAs was reached. In total, 14,354 sgRNAs targeting 748 genes were selected. The addition of 646 non-targeting control sgRNAs brought the total guide count for RLC14 to 15,000.

Combining RLC3 (16,681 sgRNAs) and RLC14 to make RLC25, the final library contained 31,143 sgRNAs: 30,306 sgRNAs targeting 4,005 genes, along with 837 non-targeting control sgRNAs (538 sgRNAs are shared between RLC3 and RLC14). This represents a reduction of 65,557 sgRNAs relative to the reference Mtb library RLC12 (ref. 10), while maintaining genome-wide coverage.

CRISPRi library production

GI libraries

The Msm and Mtb GI libraries were cloned using an approach based on CombiGEM20,75. This involved first constructing a library of barcoded single-sgRNA constructs containing strategically placed restriction sites. These sites enabled efficient derivation of a dual-sgRNA library containing all pairwise sgRNA permutations, with the corresponding barcodes in close proximity to one another to facilitate high-throughput sequencing.

Single-sgRNA library production

The biotechnology company Genewiz was commissioned to synthesize the single-sgRNA libraries. They received a list of the selected sgRNA sequences, each of which had been assigned a unique 18 base-pair barcode (Supplementary Table 1). Barcode sequences were derived from a published list (https://github.com/hawkjo/freebarcodes)76 after excluding those that would create restriction sites incompatible with downstream cloning. Each sgRNA–barcode pair was cloned individually into the plasmid backbone plRL389 (see ‘Plasmid construction’), with maintenance of restriction sites necessary for CombiGEM cloning. This approach prevents template switching77,78 between the sgRNA and its linked barcode, preserves sequence fidelity and ensures even representation during pooling. In the final constructs, EcoRI and SalI sites flank the sgRNA–barcode cassette, and MfeI and XhoI sites occupy a short spacer region between the sgRNA and the barcode. These restriction enzymes were selected because they recognize distinct sequences but generate compatible overhangs. The crucial region of each construct, encompassing the restriction sites, sgRNA and barcode, was verified by Genewiz using Sanger sequencing. The constructs were then pooled in equimolar ratios to generate the single-sgRNA libraries, which were designated RLGGI004A (Msm) and RLGGI003A (Mtb). After receipt in the laboratory, the barcode region of these libraries was amplified and sequenced to confirm the presence and balanced representation of all constructs (see ‘Library preparation for Illumina sequencing’).

Dual-sgRNA library production

The single-sgRNA libraries were designed so that digestion with MfeI and XhoI yields a backbone library, whereas digestion with EcoRI and SalI generates an insert library. The resulting compatible overhangs permit insertion of each sgRNA–barcode cassette between the sgRNA and barcode of every construct. In each digest, the sgRNA and its corresponding barcode remain physically linked. This cloning scheme purposefully omits amplification steps that can introduce template-switching artefacts, ensuring that barcodes accurately represent the sgRNA sequences contained within a construct. Moreover, it produces dual-sgRNA constructs in which the two barcodes are separated by only six base pairs, preserving barcode fidelity during library preparation and sequencing (Extended Data Fig. 1a).

To construct the Msm dual-sgRNA library (RLGGI004B), the single-sgRNA library RLGGI004A was first concentrated by alcohol precipitation, and 36 μg of RLGGI004A was then digested with MfeI (NEB, R3589) and XhoI (NEB, R0146) to obtain a library of backbones. An additional 36 μg of RLGGI004A was digested with EcoRI (NEB, R3101) and SalI (NEB, R3138) to obtain a library of inserts. The backbone and insert fragments were gel-purified (QIAGEN, 28706) and spot-dialysed (Millipore, VSWP02500) on 0.025-μm MCE membranes. The purified fragments were ligated overnight at 16 °C using T4 DNA ligase (NEB, M0202) in 28 parallel 50-μl reactions, each containing 75 fmol of backbone and insert. After heat inactivation (65 °C for 10 min), ligation products were pooled and spot-dialysed, generating the final dual-sgRNA library for transformation into E. coli.

An identical procedure was used to construct the Mtb dual-sgRNA library (RLGGI003B) from the single-sgRNA library RLGGI003A, with the exception that eight ligation reactions yielded sufficient material for E. coli library transformations.

RLC25

As described in ‘CRISPRi library design’, RLC25 consists of two sub-libraries: RLC3 and RLC14. The cloning of RLC3 has been described previously10.

To clone RLC14, an sgRNA oligonucleotide library was synthesized by Agilent Technologies (SureGuide Custom CRISPR Guide Library; G7555B), and cloning was performed as described for RLC3, with the following modifications:

  1. (1)

    Fifteen 50-µl PCR reactions were performed to amplify the oligonucleotide library, with slight adjustments to the PCR cycling conditions: 98 °C for 30 s; 11 cycles of 98 °C for 10 s, 60 °C for 10 s, 72 °C for 15 s; and 72 °C for 120 s.

  2. (2)

    The plasmid plRL58 was used in place of plRL2 as the CRISPRi backbone. Unlike plRL2, plRL58 lacks the L5 integrase, which must be supplied in trans during transformation into mycobacteria. The amplified oligonucleotide library was cloned into BsmBI-digested plRL58 using 16 × 20-µl Golden Gate reactions.

E. coli library transformations

GI libraries

The RLGGI004B and RLGGI003B ligation products were transformed into MegaX DH10B T1R Electrocomp Cells (Invitrogen, C640003) for amplification.

RLGGI004B

Twenty-five tubes of MegaX cells (100 µl each) were thawed on ice. To each tube, 75 ng (41 µl) of ligation product was added, followed by an equal volume (141 µl) of ice-cold 10% glycerol. Each tube was divided into three aliquots for transformation, resulting in a total of 75 transformations. For each transformation, 94 µl of the cell, DNA and glycerol mixture was transferred to a 0.1-cm electroporation cuvette (Bio-Rad, 1652089) and electroporated using the following settings: 2,000 V, 200 Ω and 25 µF. Immediately after electroporation, 2 × 300 µl of SOC medium was added to the cuvette, mixed by pipetting and transferred to a recovery vessel pre-loaded with SOC medium at a volume proportional to the number of transformations to be pooled in that vessel (1 ml per transformation). The total culture volume in each recovery vessel was maintained at approximately one-third of the vessel’s capacity. Recovery cultures were incubated at 37 °C with shaking for 1.5 h, then pooled into a single container. The pooled recovery culture was diluted to an optical density at 600 nm (OD600 nm) of 0.2 in LB with 50 µg ml−1 kanamycin for selection79 and distributed across six 2-l flasks to maintain adequate headspace. After overnight growth (37 °C with shaking), the plasmid library was isolated using a QIAGEN Plasmid Giga Kit (QIAGEN, 12191). Finally, the quality of the library—defined by the presence and balanced representation of expected barcode pairs—was confirmed by deep sequencing (see ‘Library preparation for Illumina sequencing’).

RLGGI003B

A similar procedure was used, with the following exceptions:

  1. (1)

    Fewer transformations (30) were performed, requiring ten tubes of MegaX cells. The DNA-to-cell ratio was maintained at 75 ng per 100 µl, necessitating the addition of 24 µl of ligation product and 124 µl of 10% glycerol per tube, with transformations carried out in 83-µl aliquots.

  2. (2)

    Library selection and outgrowth were performed on solid medium rather than in liquid culture. The pooled recovery culture (around 50 ml in total) was spread at 1 ml per plate onto pre-warmed LB agar supplemented with 50 µg ml−1 kanamycin in 50 bioassay dishes (245 mm × 245 mm; Corning, 431110). The plates were incubated at 37 °C for 18 h, after which transformants were collected by scraping.

RLC25

The transformation of sub-library RLC3 into MegaX DH10B T1R Electrocomp Cells has been described previously10. Sub-library RLC14 was transformed in the same manner, using 16 individual transformations. For each transformation, 2 µg of the dialysed Golden Gate product (54.2 µl) was mixed with 400 µl cells. Recovered cells from all transformations were pooled and plated across 16 bioassay dishes.

Mycobacterial library transformations

GI libraries

Validated dual-sgRNA libraries were transformed into mycobacterial strains already equipped with the invariant elements of the CRISPRi system (see ‘Plasmid construction’). In preparation for library transformations, these strains were rendered electrocompetent using standard procedures. In brief, cultures were grown to late log phase (OD600 nm ≈ 1) and treated with 200 mM glycine for one doubling time. Cells were then collected by centrifugation, washed three times with 10% glycerol and resuspended to obtain a 100-fold concentrated cell suspension.

Msm

Ninety-three transformations were performed to generate the Msm dual-sgRNA library. For each transformation, 1 µg of RLGGI004B plasmid DNA was added to 100 µl electrocompetent cells. The cell and DNA mixture was transferred to a 0.2-cm electroporation cuvette (Bio-Rad, 1652086) and electroporated using the following settings: 2,500 V, 700 Ω and 25 µF. Each transformation was recovered in 3 ml of appropriately supplemented 7H9 medium (see ‘Mycobacterial cultures’). Up to 17 transformations were pooled, yielding six recovery cultures of 51 ml or less in 250-ml Erlenmeyer flasks (Corning, 430183). Recovery cultures were incubated at 37 °C with shaking for three hours (approximately one doubling time) and subsequently combined in a single vessel. For selection and outgrowth of transformants, the pooled recovery culture was diluted to an OD600 nm of 0.2 in 7H9 medium containing 5 µg ml−1 zeocin and 20 µg ml−1 kanamycin, and distributed across 14 × 500-ml PETG media bottles (Thermo Fisher Scientific, 09-923-16B) to maintain adequate headspace. After reaching an OD600 nm of about 3, outgrowth cultures were pooled and used to inoculate 10 × 50-ml cultures at a density sufficient to preserve library coverage and avoid bottlenecking. After ten generations of serial passaging, these cultures were pooled and passed through a 10-μm cell strainer (pluriSelect, 437001003) to yield the final de-clumped library. A portion of the library was stored in glycerol stocks at –80 °C. A 10 OD600 nm unit (ODU) sample was taken for genomic DNA (gDNA) extraction and confirmation of library quality by deep sequencing (see ‘Genomic DNA extraction’ and ‘Library preparation for Illumina sequencing’).

Mtb

The Mtb dual-sgRNA library was generated twice in independent rounds of transformation. In the first round, 100 transformations were performed. For each transformation, 1 µg of RLGGI003B plasmid DNA was added to 200 µL of electrocompetent cells, and the mixture was electroporated as described above. Each transformation was recovered in 5 mL of appropriately supplemented 7H9 medium (see ‘Mycobacterial cultures’). Ten transformations were pooled, yielding 10 × 50-ml recovery cultures in T-225 vented tissue culture flasks (Corning, CLS431082). Recovery cultures were incubated at 37 °C for 24 h (approximately one doubling time), after which they were centrifuged to pellet the cells, pooled and resuspended in 70 ml 7H9 medium. For selection and outgrowth of transformants, the pooled recovery culture was spread at 0.7 ml per plate onto 7H10 agar supplemented with 20 µg ml−1 zeocin and 20 µg ml−1 kanamycin in 100 bioassay dishes (245 mm × 245 mm; Corning, 431110). The plates were incubated at 37 °C for 22 days, after which transformants were collected by scraping. Scraped cells were resuspended in 25% glycerol in gentleMACS M tubes (Miltenyi Biotec, 130093236) and homogenized for 90 s on a gentleMACS Octo Dissociator (Miltenyi Biotec, 130095937). A portion of the pooled homogenate was stored in glycerol stocks, and the remainder was used to inoculate 10 × 50-ml cultures for serial passaging. During serial passaging, all cultures developed a fast-growing, carbenicillin-sensitive bacterial contaminant, which had been present at very low abundance in the glycerol stocks but ultimately outcompeted Mtb, presumably owing to a shorter doubling time. Because Mtb is naturally resistant to carbenicillin, two new 50-ml cultures were initiated from glycerol stocks and passaged for ten generations in the presence of 100 µg ml−1 carbenicillin. At each passage, aliquots were plated on blood agar, and the absence of growth on these plates confirmed successful elimination of the contaminant. The cured cultures were pooled and passed through a 10-μm cell strainer to obtain the final de-clumped library. A portion of the library was stored in glycerol stocks, and a sample was taken for gDNA extraction and library quality assessment by deep sequencing (see ‘Genomic DNA extraction’ and ‘Library preparation for Illumina sequencing’). To minimize the potential effects of carbenicillin exposure, the library was passaged for an additional ten generations without carbenicillin before starting Mtb experiment 1.

Separately, a second round of 50 transformations was performed using the same electroporation and recovery procedures as in the first round, yielding 5 × 50-ml recovery cultures. After recovery, selection and outgrowth were performed in liquid culture instead of on plates, following a protocol similar to that used for Msm. Recovery cultures were combined in a single vessel, diluted to an OD600 nm of 0.2 in 7H9 medium containing 20 µg ml−1 zeocin and 20 µg ml−1 kanamycin and distributed across 50 T-225 vented tissue culture flasks to maintain adequate aeration. After reaching an OD600 nm of about 1.2, outgrowth cultures were pooled and used to inoculate 10 × 50-ml cultures for serial passaging and de-clumping as described above. A portion of the de-clumped library was stored in glycerol stocks, and a sample was taken for gDNA extraction and library quality assessment by deep sequencing (see ‘Genomic DNA extraction’ and ‘Library preparation for Illumina sequencing’). This library was used for Mtb experiment 2.

RLC25

The transformation of sub-library RLC3 into wild-type Mtb has been described previously10. Sub-library RLC14 was transformed in the same manner, with the exception that because of the absence of L5 integrase, for each transformation, 1 µg of RLC14 plasmid DNA was combined with 1 µg of plRL19 to permit integration of library constructs into the mycobacterial chromosome.

To generate RLC25, one 1-ml aliquot (one OD580 nm unit) of the final homogenized RLC3 library and one 1-ml aliquot (one OD580 nm unit) of the final homogenized RLC14 library were thawed, and each aliquot was inoculated into 24 ml 7H9 medium supplemented with kanamycin (10 μg ml−1) in separate 75-cm2 vented tissue culture flasks (starting OD580 nm ≈ 0.04). The cultures were incubated for six days at 37 °C up to an OD580 nm of about 2, passed through a 10-µm cell strainer to obtain a single-cell suspension and pooled in a 1:0.86 ratio corresponding to relative library size. The OD580 nm of the mixture was adjusted to 1.0, and aliquots were frozen at −80 °C. Genomic DNA was extracted from the library pool, and library quality was confirmed by deep sequencing (see ‘Genomic DNA extraction’ and ‘Library preparation for Illumina sequencing’).

Pooled CRISPRi screens

GI screens

Pooled CRISPRi screens were performed as described previously10, incorporating two key modifications—larger culture volumes and increased passaging frequency—to avoid bottlenecking the higher-complexity GI libraries.

Msm

RLGGI004B glycerol stocks totalling about 8 ODU (around 2.4 × 109 cells; around 1,020× library coverage) were thawed into appropriately supplemented 7H9 medium (see ‘Mycobacterial cultures’) containing zeocin (5 µg ml−1) and kanamycin (20 µg ml−1). The starter culture was expanded to enable (1) collection of six 10-ODU samples (around 1,275× library coverage) for gDNA extraction and (2) inoculation of six 50-ml ‘generation zero’ cultures at an OD600 nm of 0.3 (around 1,912× library coverage) in 250-ml Erlenmeyer flasks. Of the six cultures, three replicates received ATc (100 ng ml−1) and three served as untreated controls. Time points occurred every 2 generations through 20 generations. Generations were calculated empirically using the optical density of the cultures at each time point. Because the libraries target essential genes, +ATc cultures exhibit a growth defect relative to –ATc cultures. To account for this difference in doubling time, +ATc and –ATc cultures were collected at different times so that they differed by no more than 0.5 generations at the time of collection. At each time point, 10-ODU samples were collected for gDNA extraction, and cultures were back-diluted to an OD600 nm of 0.3 with ATc replenishment, maintaining CRISPRi efficacy and a tight OD600 nm range defined by library coverage requirements (lower bound) and sustained log-phase growth (upper bound).

Mtb

Both Mtb screens were performed similarly to the Msm screen, with shared modifications including growth in T-225 vented tissue culture flasks, use of 7H9 medium containing OADC instead of ADC (see ‘Mycobacterial cultures’) and a higher concentration of zeocin (20 µg ml−1). Screen-specific differences were as follows. Experiment 1 used 100 ng ml−1 ATc and was initiated from RLGGI003B glycerol stocks with similar attention to library coverage considerations as described for the Msm screen. Time points occurred every 2.5 generations through 30 generations, with back-dilution to a minimal OD600 nm of 0.2 (around 1,346× library coverage). Up to generation 10, samples from every 2.5 generations were sent for sequencing. From generation 10 to generation 30, samples from every 5 generations were sent for sequencing. Experiment 2 used 500 ng ml−1 ATc and was initiated directly from the pooled library filtrate. Time points matched those used in Msm, with back-dilution to a minimal OD600 nm of 0.25 (around 1,682× library coverage).

CGI screens

CRISPRi drug profiling screens were initiated by thawing two 1-ml aliquots of RLC25. Each aliquot (one OD580 nm unit) was inoculated into 24 ml 7H9 supplemented with kanamycin (10 μg ml−1) in separate 75-cm2 vented tissue culture flasks (starting OD580 nm ≈ 0.04). Cultures were expanded for five days to an OD580 nm of about 1.2, then pooled and passed through a 10-μm cell strainer to obtain a single-cell suspension. The single-cell suspension was diluted to an OD580 nm of 0.5 in a total volume of 75 ml 7H9; ATc was added at a concentration of 100 ng ml−1 to initiate target pre-depletion; and the culture was split across three 75-cm2 flasks. After 24 h of incubation, the pre-depletion cultures were pooled and used to inoculate triplicate cultures at an OD580 nm of 0.05 in 3.3 ml 7H9 medium, supplemented with ATc (100 ng ml−1), kanamycin (10 μg ml−1) and the compound under investigation (dissolved in dimethyl sulfoxide; DMSO). Drug concentrations ranged from 0.03125-fold to twofold the pre-determined minimum inhibitory concentration yielding 90% growth inhibition (MIC90), for a total of nine drug concentrations alongside the no-drug (DMSO-only) control. Cultures were grown in 12.5-cm2 vented tissue culture flasks for 14 days at 37 °C with 5% CO2 with replenishment of ATc on day 7. After 14 days, OD580 nm values were measured to determine the experimental MIC90 for each compound, and all cultures were collected for gDNA extraction.

Genomic DNA extraction

GI experiments

Ten ODU of culture were centrifuged at 4,000 rpm and washed with phosphate-buffered saline (PBS) containing 0.05% Tween-80. Cell pellets were resuspended in a 1:1 mixture of TE buffer (10 mM Tris and 1 mM EDTA, pH 8.0) and phenol:chloroform:isoamyl alcohol (25:24:1; PCI), transferred to tubes prefilled with 0.1-mm silica beads (MP Biomedicals, 1169110-CF) and mechanically lysed using two 30-s cycles at 10,000 rpm on a Precellys Evolution homogenizer. After phase separation by centrifugation, the upper aqueous phase containing gDNA was transferred to a new tube. For Mtb-derived samples, a second 1:1 PCI extraction was performed in the presence of 500 mM NaCl, with gentle mixing followed by 30-min incubation at room temperature. The aqueous phase—either singly (Msm) or dually (Mtb) extracted—was treated with RNase A at 37 °C for 45 min to degrade RNA. A final 1:1 chloroform extraction was performed to remove residual phenol, after which gDNA was recovered by alcohol precipitation. To each sample, 1/10 volume of 3 M sodium acetate (pH 5.2) and 1/300 volume of GlycoBlue (Thermo Fisher Scientific, AM9516) were added, followed by either 1 volume of isopropanol (Mtb) or 2.5 volumes of ethanol (Msm). Samples were incubated overnight at –20 °C, then centrifuged to pellet the DNA. The pellet was washed twice with 70% ethanol and resuspended in either molecular-biology-grade water or elution buffer (QIAGEN, 19086). DNA concentration was measured with a Qubit 2.0 Fluorometer (Invitrogen) to ensure accurate input into library amplification reactions.

CGI experiments

Genomic DNA was extracted using the Omega Bio-Tek Mag-Bind Universal Pathogen DNA Kit (M4029), along with a Geno/Grinder for cell lysis and a Hamilton Starlet for partial automation. The 3.3-ml cultures were transferred to 2.2-ml 96-well deep-well plates, with 2 ml of each culture allocated to one plate for extraction and the remaining volume to a second plate as a back-up. All plates were centrifuged (4,000 rpm, 10 min) to pellet the bacteria, and the supernatant was discarded. Back-up plates were covered with foil seals, bagged and stored at –80 °C. For the extraction plates, each pellet was resuspended in 520 μl SLX-Mlus buffer, and foil seals were applied. Plates were incubated at 80 °C for 1.5 h to achieve heat inactivation, allowing removal from the BSL-3 facility. After brief centrifugation (4,000 rpm, 1 min) to facilitate clean detachment of the foil seals, plate contents were transferred to an EZ-96 disruptor plate and resealed for mechanical lysis on a Geno/Grinder (2 × 1-min cycles at 1,750 rpm). After centrifugation (2,000g, 1 min) and unsealing, 53 μl DS buffer and 20 μl proteinase K were added to each well. Plates were resealed, vortexed and incubated at 70 °C for 15 min, with an additional vortexing step at the halfway point. Plates were again centrifuged (4,000 rpm, 10 min) and unsealed, and the remainder of the extraction process was done using a Hamilton Starlet liquid handling system. During this process, 300 µl was transferred from each well of the disruptor plate to the corresponding well of a 1.7-ml 96-well deep-well plate. To each well, 300 µl RBB buffer, 300 µl XP2 binding buffer and 20 µl Mag-Bind Particles RQ were added. Plates were vortexed at 1,000 rpm and incubated at room temperature for 10 min. Each plate was then placed on a magnetic separation device, and after 5 min the supernatant was aspirated and discarded. Plates were removed from the magnetic device, and 600 µl VHB buffer was added to each sample. Plates were vortexed, incubated at room temperature for 2 min and returned to the magnetic device. The supernatant was removed, and the VHB wash was repeated once. This procedure was then repeated using SPM buffer. After removal of the final supernatant, the magnetic beads were allowed to dry on the magnetic device at room temperature for 10 min. Subsequently, 50 µl of EB buffer was added to each well, plates were vortexed and samples were heated to 70 °C for 5 min on a heating station. Plates were then returned to the magnetic device, and 50 µl of gDNA from each well was transferred into the corresponding well of a clean 96-well PCR microplate. DNA concentration was measured with a Thermo Fisher Scientific NanoDrop ND-8000 spectrophotometer.

Library preparation for Illumina sequencing

GI libraries

Illumina libraries were constructed as described previously10, with two key modifications—increased gDNA input and the use of custom amplification and sequencing primers—to accommodate the increased size and unique design of GI libraries. The barcode region was amplified from gDNA using NEBNext Ultra II Q5 Master Mix (NEB, M0544). The two barcodes are separated by only 6 bp, which minimizes template switching during amplification for deep sequencing. For the Msm screen, as well as for Mtb experiment 2, 4 µg of gDNA input was used, corresponding to around 5.6 × 108 genome copies and around 238× library coverage for Msm, and around 8.9 × 108 genome copies and around 398× library coverage for Mtb. For Mtb experiment 1, 2 µg of gDNA input was used, corresponding to around 199× library coverage. PCR cycling conditions were: 98 °C for 45 s; 12 cycles of 98 °C for 10 s, 64 °C for 30 s and 65 °C for 20 s; 65 °C for 5 min. Each PCR reaction included a common forward primer (0.5 μM final concentration) and a unique indexed reverse primer (0.5 μM). The forward primer contained a P5 flow cell attachment sequence. The reverse primer contained a P7 flow cell attachment sequence, a unique index for multiplexing and a standard Illumina read 2 sequencing primer binding site. After PCR amplification, each approximately 350-bp amplicon was purified using sparQ PureMag beads (Quantabio, 95196) using two-sided selection (0.75× and 0.12×). Eluted amplicons were quantified with a Qubit 2.0 Fluorometer, and amplicon size and purity were quality-controlled by visualization on an Agilent 4200 TapeStation (Agilent Technologies G2991AA, 5067-5585 and 5067-5584). To assess library quality and ensure equal read depth across samples, individual amplicons were at first pooled in equal volumes and sequenced on an Illumina MiSeq according to the manufacturer’s instructions. To increase sequence diversity, a PhiX spike-in of 10% (PhiX Control v3; Illumina, FC-110-3001) was included. Sequencing was performed in single-read mode (1 × 50 bp) with an 8-bp i7 index read. A custom sequencing primer binding immediately adjacent to the barcode cassette was spiked into the standard Illumina read 1 primer mix to enable sequencing of both sample barcodes and PhiX DNA. After the MiSeq run, amplicons were repooled such that the volume of each sample was scaled according to its read count. The rebalanced pool was then sequenced on an Illumina NovaSeqX Plus sequencer according to the manufacturer’s instructions, using the same sequencing conditions as in the MiSeq run. Library amplification primers and the custom sequencing primer are listed in Supplementary Table 6.

Amplicon libraries from plasmid DNA were prepared using the same protocol as for gDNA, except that 10 ng DNA input and 11 PCR cycles were used. For validation of single-sgRNA libraries, sequencing was performed using the Illumina MiSeq Nano platform with 25-bp reads (sufficient because of the lower library complexity and single-barcode design).

RLC25

Illumina libraries were constructed as described previously39.

Quantification and statistical analysis

Quantification of sgRNA depletion

Sequencing reads were aligned to a reference set of expected sequences using subread-align (v.1.6.0) with default parameters80. For GI libraries, the reference set included barcode pairs, and read counts were subsequently assigned to the corresponding sgRNA pairs. For RLC25, the reference set included the sgRNA sequences themselves. Differences in sequencing depth between samples were normalized using the trimmed total reads (TTR) method73.

Depletion of sgRNA pairs in GI screens was quantified by estimating the L2FC of the mean counts in the +ATc samples relative to the mean counts in the –ATc samples. The pairs containing two non-targeting control sgRNAs were used in a second normalization step that sets their median L2FC as equal to 0, controlling for the enrichment of non-targeting guides in the library10. Finally, L2FC data points were subjected to quality-control filters, including removal of points that flatlined10 or reached the limit of detection (fewer than five mean observed reads). sgRNA pairs with fewer than three data points passing quality control were excluded from further analysis.

Depletion of sgRNAs in chemical–genetic screens was quantified by MAGeCK81 (v.0.5.7), using the alphamedian method and normalizing counts relative to non-targeting controls (using the flags ‘--gene-lfc-method alphamedian --norm-method control’).

Modelling sgRNA depletion over time

The depletion of sgRNA pairs in GI experiments was modelled using a previously established Bayesian piecewise linear regression model, referred to as a two-line model10. In brief, the parameters αl and βl represent the intercept and depletion of the initial phase of depletion, which then transitions at a point γ, to a second phase with a rate of depletion βe. With this model, a prediction of L2FC at a generation, G, can be obtained as follows:

$$\text{L2FC}(G)={\alpha }_{{\rm{l}}}+{\beta }_{{\rm{l}}}G\,\,\,\,\,{\rm{If}}\,G\le \gamma $$

$$\text{L2FC}(G)=({\alpha }_{{\rm{l}}}+{\beta }_{{\rm{l}}}\gamma )+({\beta }_{{\rm{e}}}(G-\gamma ))\,\,\,\,{\rm{If}}\,G > \gamma $$

Quantification of GIs

GI score

Samples from the Bayesian two-line model were obtained for all sgRNA pairs using Stan (cmdstan v.2.35, cmdstanpy v.1.2.4)82. To quantify a GI score, we first defined the degree to which pairs of sgRNAs deviated from expectations using an additive model of L2FC (which is equivalent to a multiplicative model in linear space)83. Single-targeting pairs, in which a targeting sgRNA is paired with a non-targeting control, are used to determine the fitness effect of knocking down a particular gene. Expected effects of knocking down two genes are therefore taken as the sum of individual effects:

$$E[{g}_{1}=a,{g}_{2}=b]=\text{L2FC}(G|{g}_{1}=a,{g}_{2}=\text{NT})+\text{L2FC}(G|{g}_{1}=\text{NT},{g}_{2}=b),$$

where \(\text{L2FC}(G|{g}_{1}=a,{g}_{2}=b)\) represents the L2FC predicted by the Bayesian two-line model at generation G for a given pair of guides, a and b, at sgRNA positions 1 and 2, respectively, and NT represents the non-targeting controls. Deviations from expectations are then defined as follows:

$$\Delta \text{L2FC}(G|{g}_{1}=a,{g}_{2}=b)=\text{L2FC}(G|{g}_{1}=a,{g}_{2}=b)-E[{g}_{1}=a,{g}_{2}=b].$$

To summarize the deviations to a single value, we define ΔL2FC as the deviation from expectations at G = 25.

To address the possibility of a floor effect84 (cases in which the expected phenotype exceeds the maximal fitness cost that can be accurately detected in our experiment), a generalized additive model (GAM)85 was fitted using the bam function from the mgcv package (v.1.9-3) in R (v.4.3.1) to correct the observed deviations. Deviations were modelled as:

$$\Delta \text{L2FC}({g}_{1}=a,{g}_{2}=b) \sim \hat{s}(E[{g}_{1}=a,{g}_{2}=b]),$$

where \(\hat{s}\) is the penalized regression spline GAM (with k = 20, method = “ML”, scaled t-distribution family). By modelling deviations in this manner, the penalized smoothing encourages predictions to follow the observed trend in data-sparse regions (that is, a near-linear continuation of the observed versus expected relationship), effectively reducing abrupt changes in deviations when there is little data to support them. The final, corrected, deviations were thus calculated as:

$$\begin{array}{c}{\Delta }^{{\prime} }({g}_{1}=a,{g}_{2}=b)=\text{L2FC}({g}_{1}=a,{g}_{2}=b)\\ -(E[{g}_{1}=a,{g}_{2}=\,b]+\hat{s}(E[{g}_{1}=a,{g}_{2}=b])).\end{array}$$

Finally, the GI score for a pair of genes A and B was summarized as the median of the Δ′ for all of the guide pairs targeting those genes; that is,

$$\text{GI}(A,B)=\text{median}({\Delta }^{{\prime} }(A,B)).$$

Statistical significance

A Bayesian mixture model was used to identify statistically significant interactions. The mixture consisted of two distributions: a normal distribution representing the null hypothesis of no interaction with a prior mean of 0, and a uniform distribution representing the alternative hypothesis. The posterior probability of a GI score belonging to the alternative hypothesis was calculated, and interactions were classified as likely interactions if the posterior probability was greater than 0.5, and as confident interactions if the posterior probability was greater than 0.95. We use a cut-off of 0.5 for likely interactions because this corresponds to the point at which GI scores are more likely to be explained by the uniform distribution than by the normal distribution.

Joint GI model

To integrate the evidence from the two independent Mtb screens (experiment 1 and experiment 2) into a single, directly interpretable probability, a joint Bayesian mixture model was fitted to the paired GI scores. For each gene pair, the GI score in each experiment and its standard error (that is, mean GI score and the standard deviation of Δ′ from the Bayesian two-line model) were taken as the x,y coordinates and their measurement uncertainties.

The mixture consisted of five components defined over the range of GI scores: a null component and four interaction components, one for each quadrant. The null hypothesis of no interaction was represented by a bivariate normal distribution centred near the origin, with a correlation parameter, ρ, given a weakly informative prior centred at zero. Allowing ρ to be a free parameter allowed the model to learn any existing correlation observed in the null distribution, owing to each screen using the same library and protocols. Per-pair measurement error was incorporated into the null by adding each pair’s squared standard error to the variance on the corresponding axis and scaling ρ accordingly (see Stan model code for implementation details). This approach incorporates the uncertainty associated with each GI-score estimate, down-weighting imprecise measurements while integrating both the consistency and the precision of observed effects across the two screens.

The four interaction components corresponded to the concordant negative quadrant, the concordant positive quadrant and the two discordant quadrants in which the two experiments disagreed in sign (one positive and the other negative). Each quadrant was modelled as a product of per-axis uniform densities over the range of observed scores, which were winsorized at the 0.1st and 99.9th percentiles to reduce the influence of outliers. Because pairs near a quadrant edge could belong to a neighbouring quadrant owing to measurement uncertainty, the uniform densities were convolved with the per-pair measurement error at the origin. Thus, a pair lying near an axis is assigned partly to each quadrant, and the model identifies the most likely membership from the mixture. The mixture weights were given a Dirichlet(2, 2, 36) prior over the concordant, discordant and null components, placing an 18-fold higher weight on the null, reflecting the expectation that most gene pairs do not interact. Concordant and discordant weights were divided evenly between their two respective quadrants (two discordant quadrants and the concordant positive and concordant negative quadrants). The model was fitted using Stan (cmdstan v.2.35, cmdstanpy v.1.2.4). For each gene pair, the posterior probability of belonging to each component was computed and summarized as the probability of a negative, positive or discordant interaction, and of no interaction. As in the single-experiment analysis, a pair was classified as a likely interaction when the largest of its interaction-class probabilities exceeded 0.5 and as a confident interaction when it exceeded 0.95, and was labelled negative, positive or discordant according to the dominant class.

Hierarchical clustering analysis

To explore patterns of similarity in GI profiles, the GI matrix was hierarchically clustered using average linkage with Pearson correlation as the similarity measure, as implemented in SciPy86. At each internal node of the clustering hierarchy, the descendant genes were tested for Gene Ontology (GO) enrichment using GOATOOLS87. GO annotations were obtained from UniProt and mapped to the go-basic ontology, with the full set of genes in the GI map used as the background. Enrichment P values were calculated using Fisher’s exact test and corrected for multiple-hypothesis testing using the Benjamini–Hochberg procedure. Nodes were designated as key nodes when they represented the most significant occurrence (false discovery rate (FDR) < 0.005) for at least one biological process term. Each key node was labelled with its most significant GO term, and a GO-DAG specificity constraint was enforced such that no parent node carried a more specific term than its descendants. To reduce redundancy, key nodes with highly overlapping parent gene sets (Jaccard ≥ 0.8) were removed. Distinct colours were assigned to the remaining key nodes and propagated through their descendant branches, producing functionally labelled dendrograms in which leaves inherited the most specific term in their lineage.

Pathway enrichment

Pathways were defined using a combination of KEGG and PATRIC pathways together with manual curation (Supplementary Table 4). Genes were assigned to pathways on the basis of manual curation and AI-assisted literature review (Claude Sonnet 4.5). Gene pairs were categorized as having a significant GI (p > 0.5) or not, and Fisher’s exact test was used to assess the association between pathway pairs and significant GIs. Self-interactions, in which a gene is targeted twice (that is, A + A), were excluded from the analysis. The resulting P values were adjusted for multiple comparisons using the Benjamini–Hochberg procedure. The analysis was performed twice: once using all gene pairs and once excluding potential polar effects (defined as pairs in which the genes were five genes or fewer apart), to assess how significance changed.

Co-folding

Sequences for selected genes were co-folded with AlphaFold-Multimer28 (v.3) to predict protein complex structures and AlphaFold34 to predict PPIs. All model inputs for Mtb used amino acid sequences from the H37Rv reference proteome (UniProt: UP000001584). All model inputs for Msm used amino acid sequences translated using BioPython from strain mc2155 (RefSeq: NC_008596.1), with the following exceptions. MSMEG5414 in the reference genome was split into proteins MSMEG5414A and MSMEG5414B. Four genes are misannotated in the reference genome as pseudogenes, and their translations were manually included: MSMEG4629:folC and MSMEG1828 from another genome (GenBank: CP009494); MSMEG1843:ahcY from UniProt (accession A4ZHR8); and MSMEG1400:fusA1 from a previously reported sequence88,89. All sequence inputs for Msm can be found in the file msm_essential_proteome.fa in our linked GitHub repository (see ‘Code availability’). AlphaFold-Multimer inputs were formatted as pairs of monomers28. These complexes were predicted using LocalColabFold90 v.1.5.5 (https://github.com/YoshitakaMo/localcolabfold)90 on an NVIDIA L40S GPU and 16 24-GB CPUs. The model was run with three recycles and AMBER relaxation. AlphaFold3 inputs were prepared following a modified pooled AlphaFold3 approach27. Within the bin-packing strategy, smaller protein sequences were prioritized to maximize the number of monomers per pool. In both cases, five predicted structures were generated per input. Predicted TM scores (pTMs) and interface predicted TM scores (ipTMs) for any given gene or gene pair, respectively, were taken as a mean across all models. ipTMs were corrected for combined protein length as previously described27. In Supplementary Tables 2 and 3, columns from pooled inputs on the web server have the suffix or infix ‘_webserver’.

Structures and scores were also generated locally for Mtb using the same pooled inputs as the AlphaFold3 web server. Per-chain multiple sequence alignments were generated with ColabFold v.1.5.590 using MMseqs2 18.8cc5c91 against the UniRef30 (2023-02) and ColabFold environmental clusters (2021-08; comprising filtered BFD, MGnify, MetaEuk and SMAG sub-databases), with GPU-accelerated prefiltering (--gpu 1 --prefilter-mode 1). The output was written directly in AlphaFold3-compatible JSON form (--af3-json), yielding one unpaired multiple sequence alignment per chain. To avoid recomputing identical alignments across the many overlapping pools in this study, sequences were keyed by SHA-256 hash and computed only once; an indexed central cache of all previously generated alignments (about 3,800 unique sequences) was consulted on every job, with cache hits delivering byte-identical multiple sequence alignments and cache misses triggering a fresh ColabFold run that subsequently updated the cache. Co-folding was performed with AlphaFold3 v.3.0.14 executed inside an Apptainer/Singularity container on NVIDIA H100 80GB HBM3 GPUs. For each pool, the per-chain multiple sequence alignments were combined into a single AlphaFold3 input JSON (dialect alphafold3, version 1), populating unpairedMsa from the ColabFold output; pairedMsa was left empty, deferring to AlphaFold3’s internal pairing logic. AlphaFold3 was invoked via run_alphafold.py with --flash_attention_implementation=triton and --resolve_msa_overlaps=false; all other parameters retained their AlphaFold3 defaults. A single random seed (modelSeeds = [1]) was used per input, producing five diffusion-sampled structures per pool from which downstream pTM and ipTM statistics were derived as described above. In Supplementary Table 3, the resulting columns have the suffix or infix ‘_local’.

For both the local and the web-server results, the total number of predicted PPIs for each single protein is listed with the prefix ‘degree’, following notation from graph theory. Each of these columns is suffixed with 1 or 2 to indicate the monomer to which it corresponds. These degrees are calculated cumulatively at three ipTM cut-offs (see Fig. 3f): for example, ‘degree_low_webserver1’ refers to all PPIs predicted for gene 1 with low or higher confidence using the web server.

Minimum inhibitory concentration assay

All compounds were dissolved in DMSO (VWR, V0231) and dispensed using an HP D300e digital dispenser in a 384-well plate format using a twofold dilution series. DMSO did not exceed 1% of the final culture volume and was maintained at the same concentration across all samples. Cultures were synchronized in late logarithmic phase (OD600 nm = 0.2 to 1) and then back-diluted to a starting OD600 nm of 0.01 or 0.05. Fifty microlitres of cell suspension was plated in triplicate in wells containing the test compound. Plates were incubated standing at 37 °C with 5% CO2. The OD600 nm was evaluated using a Tecan Spark or Molecular Devices SpectraMax M2 microplate reader 7–14 days after plating, and the percentage growth was calculated relative to the DMSO vehicle control for each strain. MIC90 and MIC50 measurements were calculated in R using the drm function from the drc package92, with a four-parameter log-logistic model (LL.4). For all MIC curves, data are mean ± s.e.m. for triplicates.

Checkerboard synergy assay

Compounds were dispensed in a 384-well plate as described above for the MIC assays. For each pair of drugs, all combinations of the two-dilution series were plated, including dilution series in which each drug is present alone and a DMSO control in which neither drug is present. Each combination of concentrations was plated in triplicate. Cultures were grown and plated and the relative growth was measured as described for MIC assays, except that the cultures were seeded at OD600 nm = 0.025. Drug synergy was quantified using the synergyfinder v.3.14.0 R package93. For each checkerboard, mean percentage viability values were used to calculate the Bliss independence score94,95 at each combination of drug concentrations. The resulting scores were then scanned to find the most synergistic area (MSA) and most antagonistic area (MAA). These areas were defined as the 3 × 3 contiguous areas of the plate with the highest and lowest mean Bliss scores, respectively. The mean Bliss scores of the MSA and MAA are reported for each drug combination. Both metrics are reported to avoid bias in calling a drug pair synergistic or antagonistic. Following recommended guidelines96, drug pairs with an MSA mean Bliss score higher than 15 were considered synergistic, and drug pairs with an MAA mean Bliss score lower than –15 were considered antagonistic.

Kill curve drug synergy assay

A wild-type H37Rv culture was diluted to an OD600 nm of 0.05, and 500 μl of culture was added to each well of 24-well plates. Vertex-49 (6.25 μM or 12.5 μM) was added alone or in combination with either bedaquiline (0.45 μM or 0.89 μM) or ethambutol (3.61 μM or 7.21 μM) in triplicate for each concentration combination. Cultures were grown standing at 37 °C for 6 days and then plated in serial dilutions from 10−1 to 10−7 on 7H10 charcoal plates. Plates were grown for 21 days, and then colony-forming units (CFU) were counted. Because we did not have a full panel of concentration combinations, we opted to use the highest single agent (HSA) method of quantifying drug synergy97. In brief, the amount of killing (measured as the change in CFU per ml versus the no-drug control) with the combination of the two drugs is compared to the drug that shows the highest growth inhibition (the HSA) when used alone at the same concentration. The drug pair is considered synergistic if the combination of drugs results in more than 100-fold more killing than the HSA.

Methylene blue oxygen consumption assay

Cultures were grown for 5 days in 7H9-OADC with 100 ng ml−1 ATc to pre-deplete protein targets. The cultures were then washed twice in PBS containing 0.05% Tween-80 and adjusted to an OD600 nm of 0.15 with fresh 7H9-OADC. After dilution, 4-ml screw-cap glass vials were filled with the cultures, and methylene blue dye was added at a final concentration of 0.001% (w/v). All vials were sealed tightly using a PTFE–rubber seal (Thermo Fisher Scientific, C4015-21), wrapped with parafilm and incubated shaking at 60 rpm at 37 °C in an anaerobic pouch (Becton Dickinson GasPak EZ container systems, BD 260683) for 3–4 days, after which photos were taken to determine whether the cultures turned clear (intact respiration) or remained blue (impaired respiration).

Affinity purification and mass spectrometry

Msm growth for protein purification

A 20-ml preculture of Msm in Middlebrook 7H9 medium (Sigma) supplemented with 0.5% (w/v) bovine serum albumin, 0.08% (w/v) NaCl, 0.2% (w/v) d-glucose and 0.05% (v/v) Tween-80 was grown at 30 °C with shaking at 180 rpm to saturation. Msm with pKM444 but no other genetic modification was grown without antibiotic in the preculture medium, whereas Msm modified with a 3×Flag tag encoded in its gDNA was grown with 50 µg ml−1 hygromycin in the preculture medium. Each preculture was used to inoculate 6 l of Middlebrook 7H9 supplemented with 0.08% (w/v) NaCl, 0.2% (w/v) d-glucose, 1% (w/v) tryptone and 0.05% (v/v) Tween-80, and was grown at 30 °C with shaking at 180 rpm for 48–72 h. The cells were collected by centrifugation at 6,500g for 20 min, and pellets were resuspended in 100 ml of 50 mM Tris-HCl pH 7.5, 150 mM NaCl, 5 mM MgSO4, 5 mM benzamidine HCl, 5 mM aminocaproic acid and 1 mM PMSF. The cells were lysed with the Avestin EmulsiFlex-C3 homogenizer at 20–25 kpsi, and the lysate was clarified by centrifugation at 39,000g for 30 min, followed by isolation of the membranes by centrifugation at 200,000g for 1 h. The membranes were resuspended in TBS (50 mM Tris-HCl pH 7.5, 150 mM NaCl) and solubilized for 1 h with 1% (w/v) dodecyl maltoside. Solubilized membranes were centrifuged at 200,000g for 1 h to remove insoluble material before binding to around 1 ml anti-Flag M2 Affinity Gel (Sigma) in a plastic column. The column was washed with ten column volumes of TBS containing 0.05% (w/v) dodecyl maltoside before eluting the bound protein with three column volumes of 150 µg ml−1 3×Flag peptide in TBS containing 0.05% (w/v) dodecyl maltoside, followed by one column volume of TBS containing 0.05% (w/v) dodecyl maltoside. The eluted protein was concentrated with a centrifugal concentrator with 10-kDa molecular weight cut-off.

Preparation of protein samples for mass spectrometry

Protein (10 µg) was reduced with 20 mM dithiothreitol for 10 min at 95 °C and alkylated with 40 mM iodoacetamide for 30 min in the dark. Samples were diluted to a final concentration of 5% (w/v) SDS, and phosphoric acid was added to a final concentration of 1.2% (v/v). S-Trap protein binding buffer (165 µl) containing 90% (v/v) methanol and 100 mM triethylammonium bicarbonate (TEAB) was added to 27.5 µl acidified sample. The resulting mixture was passed through an S-trap micro column at 4,000g. The micro column was washed four times with the S-Trap protein binding buffer. Each sample was digested with 4 µg of trypsin in 20 µl of 50 mM TEAB for 1 h at 47 °C. Before elution, 4 µl of 50 mM TEAB pH 8 was added to the column. Peptides were eluted by centrifugation at 4,000g. Peptides were eluted from the column two more times with 40 µl of 0.2% (v/v) formic acid and 40 µl of 50% (v/v) acetonitrile with 0.2% (v/v) formic acid. Eluted peptides were dried and stored at –40 °C before mass-spectrometry analysis at the Network Biology Collaborative Centre at the Lunenfeld-Tanenbaum Research Institute using standard protocols for data-dependent acquisition.

Analysis of mass-spectrometry data

The SAINTexpress98,99 software package was used to score protein interactions. In brief, peptide-level counts are summed for all detected proteins for both the tagged and the untagged samples (two replicates per condition). These were then used to obtain the protein-level L2FC of counts in the tagged samples relative to the untagged samples. The SAINTexpress package uses a Bayesian model to calculate the probability of an interaction for each protein and uses these probabilities to estimate the FDR.

IMV acidification assays

IMV acidification assays were performed as previously described51. In brief, the membrane fraction from wild-type, ∆qcrB and ∆ms2764 Msm strains were isolated and resuspended in IMV resuspension buffer (50 mM Tris-HCl pH 7.5, 150 mM NaCl, 5 mM MgSO4, 5 mM 6-aminocaproic acid, 5 mM benzamidine hydrochloride and 20% [v/v] glycerol). IMV acidification experiments were performed in 96-well plates (BRANDplates pureGrade 96-well black microplates). NADH and succinate were prepared as electron donors as follows. NADH was prepared immediately before use by dissolving disodium NADH in water to a final concentration of 4 mM. Succinate was prepared by dissolving sodium succinate in water to a final concentration of 20 mM. A master solution containing ACMA assay buffer (10 mM HEPES-KOH pH 7.5, 100 mM KCl and 5 mM MgCl2), 9-amino-6-chloro-2-methoxyacridine (ACMA) dye (3.125 µM), IMVs (5 mg ml−1 total protein concentration for succinate-driven acidification; 2 mg ml−1 total protein concentration for NADH-driven acidification) and 10 µM of Q203 (with a final concentration of 2% [v/v] DMSO) or an equivalent volume of compound-free DMSO vehicle was prepared and transferred into each well of the 96-well plate. The plate was incubated in the dark for approximately 15 min at room temperature before the start of the experiment. ACMA fluorescence was measured at 25 °C with a BioTek Synergy Neo2 Hybrid Multimode plate reader (Agilent Technologies), using an excitation wavelength of 410 nm and an emission wavelength of 480 nm. IMV acidification was initiated by the addition of NADH or succinate. Fluorescence was monitored for 20 min before the addition of nigericin in ethanol to a final concentration of 2 µM with 0.02% (v/v) ethanol. Fluorescence recovery was monitored for an additional 10 min. Fluorescence recovery was calculated as the absolute difference between the fluorescence immediately before and 10 min after nigericin addition. All IMV acidification experiments were performed with at least three replicates.

Sensor spot assays

Sensor spot assays were performed as described previously100.

Cell spotting assays

Cultures were grown and passaged twice to closely match optical densities. Final cultures were diluted to an OD600 nm of 0.17 (around 50,000 cells per μl), and tenfold serial dilutions were made. Multichannel pipettes were used to spot 2 μl of culture for each dilution onto 7H10 with or without ATc, with the appropriate selection of antibiotics. Images were taken after approximately 20 generations.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

Raw sequencing data have ben deposited in the NCBI Sequence Read Archive (SRA) under project numbers PRJNA1337647, PRJNA1338999 and PRJNA1346122. Pooled AlphaFold3 co-folding used the Mtb H37Rv reference proteome (UniProt: UP000001584) and translated sequences from the Msm mc2155 reference genome (RefSeq: NC_008596.1), unless noted otherwise in the Methods. Mass-spectrometry analysis used the Msm mc2155 reference proteome (UniProt: UP000000757). sgRNA sequences were designed using the Mtb H37Rv and Msm mc2155 reference genomes (RefSeq: NC_018143.2 and RefSeq: NC_008596.1, respectively).

Code availability

References

  1. World Health Organization. Global Tuberculosis Report 2025 (WHO, 2025).

  2. Tong, A. H. Y. et al. Systematic genetic analysis with ordered arrays of yeast deletion mutants. Science 294, 2364–2368 (2001).

    Article  CAS  PubMed  Google Scholar 

  3. Costanzo, M. et al. Global genetic networks and the genotype-to-phenotype relationship. Cell 177, 85–100 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  4. Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Parsons, A. B. et al. Integration of chemical–genetic and genetic interaction data links bioactive compounds to cellular target pathways. Nat. Biotechnol. 22, 62–69 (2004).

    Article  CAS  PubMed  Google Scholar 

  6. Zhang, Y. J. et al. Global assessment of genomic regions required for growth in Mycobacterium tuberculosis. PLoS Pathog. 8, e1002946 (2012).

    Article  PubMed  PubMed Central  Google Scholar 

  7. Sassetti, C. M., Boyd, D. H. & Rubin, E. J. Genes required for mycobacterial growth defined by high density mutagenesis. Mol. Microbiol. 48, 77–84 (2003).

    Article  CAS  PubMed  Google Scholar 

  8. Sassetti, C. M. & Rubin, E. J. Genetic requirements for mycobacterial survival during infection. Proc. Natl Acad. Sci. USA 100, 12989–12994 (2003).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  9. Joshi, S. M. et al. Characterization of mycobacterial virulence genes through genetic interaction mapping. Proc. Natl Acad. Sci. USA 103, 11760–11765 (2006).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  10. Bosch, B. et al. Genome-wide gene expression tuning reveals diverse vulnerabilities of M. tuberculosis. Cell 184, 4579–4592 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  11. Qi, L. S. et al. Repurposing CRISPR as an RNA-guided platform for sequence-specific control of gene expression. Cell 152, 1173–1183 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  12. Bikard, D. et al. Programmable repression and activation of bacterial gene expression using an engineered CRISPR-Cas system. Nucleic Acids Res. 41, 7429–7437 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  13. Rock, J. M. et al. Programmable transcriptional repression in mycobacteria using an orthogonal CRISPR interference platform. Nat. Microbiol. 2, 16274 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  14. McNeil, M. B., Ryburn, H. W., Tirados, J., Cheung, C.-Y. & Cook, G. M. Multiplexed transcriptional repression identifies a network of bactericidal interactions between mycobacterial respiratory complexes. iScience 25, 103573 (2022).

    Article  CAS  PubMed  Google Scholar 

  15. Jana, B. et al. CRISPRi-TnSeq maps genome-wide interactions between essential and non-essential genes in bacteria. Nat. Microbiol. 9, 2395–2409 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  16. Poulton, N. C. et al. Beyond antibiotic resistance: the whiB7 transcription factor coordinates an adaptive response to alanine starvation in mycobacteria. Cell Chem. Biol. 31, 669–682 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  17. Dénéréaz, J. et al. Dual CRISPRi-seq for genome-wide genetic interaction studies identifies key genes involved in the pneumococcal cell cycle. Cell Syst. 16, 101408 https://doi.org/10.1016/j.cels.2025.101408 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  18. Koo, B.-M. et al. Comprehensive genetic interaction analysis of the Bacillus subtilis envelope using double-CRISPRi. Cell Syst. 16, 101406 https://doi.org/10.1016/j.cels.2025.101406 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  19. Zik, J. J. et al. Dual transposon sequencing profiles the genetic interaction landscape in bacteria. Science 389, eadt7685 (2025).

    Article  CAS  PubMed  Google Scholar 

  20. Cheng, A. A., Ding, H. & Lu, T. K. Enhanced killing of antibiotic-resistant bacteria enabled by massively parallel combinatorial genetics. Proc. Natl Acad. Sci. USA 111, 12462–12467 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  21. Lupoli, T. J., Fay, A., Adura, C., Glickman, M. S. & Nathan, C. F. Reconstitution of a Mycobacterium tuberculosis proteostasis network highlights essential cofactor interactions with chaperone DnaK. Proc. Natl Acad. Sci. USA 113, E7947–E7956 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Boutte, C. C. et al. A cytoplasmic peptidoglycan amidase homologue controls mycobacterial cell wall synthesis. eLife 5, e14590 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

  23. Fay, A. et al. Two accessory proteins govern MmpL3 mycolic acid transport in mycobacteria. mBio 10, e00850-19 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  24. Gupta, K. R. et al. An essential periplasmic protein coordinates lipid trafficking and is required for asymmetric polar growth in mycobacteria. eLife 11, e80395 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  25. Bush, M. J., Chandra, G., Bibb, M. J., Findlay, K. C. & Buttner, M. J. Genome-wide chromatin immunoprecipitation sequencing analysis shows that WhiB is a transcription factor that cocontrols its regulon with WhiA to initiate developmental cell division in Streptomyces. mBio 7, e00523-16 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

  26. Roguev, A. et al. Conservation and rewiring of functional modules revealed by an epistasis map in fission yeast. Science 322, 405–410 (2008).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  27. Todor, H. et al. Predicting the protein interaction landscape of a free-living bacterium with pooled-AlphaFold3. Mol. Syst. Biol. 22, 497–518 (2026).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  28. Evans, R. et al. Protein complex prediction with AlphaFold-Multimer. Preprint at bioRxiv https://doi.org/10.1101/2021.10.04.463034 (2022).

  29. Bailo, R. et al. The mycobacterial desaturase DesA2 is associated with mycolic acid biosynthesis. Sci. Rep. 12, 6943 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Baños-Mateos, S. et al. High-fidelity DNA replication in Mycobacterium tuberculosis relies on a trinuclear zinc center. Nat. Commun. 8, 855 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

  31. Beites, T. et al. Multiple acyl-CoA dehydrogenase deficiency kills Mycobacterium tuberculosis in vitro and during infection. Nat. Commun. 12, 6593 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  32. Jansen, R. S. et al. Aspartate aminotransferase Rv3722c governs aspartate-dependent nitrogen metabolism in Mycobacterium tuberculosis. Nat. Commun. 11, 1960 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  33. Prosser, G. A. & de Carvalho, L. P. S. Metabolomics reveal D-alanine:D-alanine ligase as the target of D-cycloserine in Mycobacterium tuberculosis. ACS Med. Chem. Lett. 4, 1233–1237 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  34. Edoo, Z. et al. BVL3572S inhibits HisC and AlaA, exploiting vitamin B6 dependency to kill Mycobacterium tuberculosis. Cell Chem. Biol. https://doi.org/10.1016/j.chembiol.2026.07.008 (2026).

    Article  PubMed  Google Scholar 

  35. Ballinger, E. et al. Opposing reactions in coenzyme A metabolism sensitize Mycobacterium tuberculosis to enzyme inhibition. Science 363, eaau8959 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  36. Abrahams, K. A. et al. Identification of KasA as the cellular target of an anti-tubercular scaffold. Nat. Commun. 7, 12581 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  37. Grzegorzewicz, A. E. et al. A common mechanism of inhibition of the Mycobacterium tuberculosis mycolic acid biosynthetic pathway by isoxyl and thiacetazone. J. Biol. Chem. 287, 38434–38441 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  38. Borrell, S. et al. Reference set of Mycobacterium tuberculosis clinical strains: a tool for research and product development. PLoS ONE 14, e0214088 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  39. Li, S. et al. CRISPRi chemical genetics and comparative genomics identify genes mediating drug potency in Mycobacterium tuberculosis. Nat. Microbiol. 7, 766–779 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  40. Krieger, I. V. et al. SuFEx-based antitubercular compound irreversibly inhibits Pks13. Nature 645, 755–763 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  41. Johnson, E. O. et al. Large-scale chemical-genetics yields new M. tuberculosis inhibitor classes. Nature 571, 72–78 (2019).

    Article  CAS  PubMed  Google Scholar 

  42. Wang, T. et al. Mtb PKNA/PKNB dual inhibition provides selectivity advantages for inhibitor design to minimize host kinase interactions. ACS Med. Chem. Lett. 8, 1224–1229 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  43. Zhang, L. et al. Structures of cell wall arabinosyltransferases with the anti-tuberculosis drug ethambutol. Science 368, 1211–1219 (2020).

    Article  CAS  PubMed  Google Scholar 

  44. Andries, K. et al. A diarylquinoline drug active on the ATP synthase of Mycobacterium tuberculosis. Science 307, 223–227 (2005).

    Article  CAS  PubMed  Google Scholar 

  45. Cacace, E. et al. Systematic analysis of drug combinations against Gram-positive bacteria. Nat. Microbiol. 8, 2196–2212 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  46. Pethe, K. et al. Discovery of Q203, a potent clinical candidate for the treatment of tuberculosis. Nat. Med. 19, 1157–1160 (2013).

    Article  CAS  PubMed  Google Scholar 

  47. Liang, Y. & Rubinstein, J. L. Structural analysis of mycobacterial electron transport chain complexes by cryoEM. Biochem. Soc. Trans. 51, 183–193 (2023).

    CAS  PubMed  Google Scholar 

  48. Brischigliaro, M. & Zeviani, M. Cytochrome c oxidase deficiency. Biochim. Biophys. Acta Bioenerg. 1862, 148335 (2021).

    Article  CAS  PubMed  Google Scholar 

  49. Hederstedt, L. Diversity of cytochrome c oxidase assembly proteins in bacteria. Microorganisms 10, 926 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  50. Lee, B. S. et al. Dual inhibition of the terminal oxidases eradicates antibiotic-tolerant Mycobacterium tuberculosis. EMBO Mol. Med. 13, e13207 (2021).

    Article  CAS  PubMed  Google Scholar 

  51. Harden, S. A., Courbon, G. M., Liang, Y., Kim, A. S. & Rubinstein, J. L. A simple assay for inhibitors of mycobacterial oxidative phosphorylation. J. Biol. Chem. 300, 105483 (2024).

    Article  CAS  PubMed  Google Scholar 

  52. Buglino, J. A. et al. Chalkophore-mediated respiratory oxidase flexibility controls M. tuberculosis virulence. eLife 14, RP105794 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  53. Theriault, M. E. et al. Iron limitation in M. tuberculosis has broad impact on central carbon metabolism. Commun. Biol. 5, 685 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  54. Herrera, N. et al. The phenotypic landscape of the mycobacterial cell. Preprint at bioRxiv https://doi.org/10.1101/2025.11.14.688347 (2026).

  55. de Wet, T. J., Winkler, K. R., Mhlanga, M., Mizrahi, V. & Warner, D. F. Arrayed CRISPRi and quantitative imaging describe the morphotypic landscape of essential mycobacterial genes. eLife 9, e60083 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  56. Yoo, R. et al. Machine learning of all Mycobacterium tuberculosis H37Rv RNA-seq data reveals a structured interplay between metabolism, stress response, and infection. mSphere 7, e0003322 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  57. Jinich, A. et al. The Mycobacterium tuberculosis transposon sequencing database (MtbTnDB): a large-scale guide to genetic conditional essentiality. Mol. Microbiol. 124, 91–101 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  58. Dunn, J. J. & Studier, F. W. T7 early RNAs and Escherichia coli ribosomal RNAs are cut from large precursor RNAs in vivo by ribonuclease 3. Proc. Natl Acad. Sci. USA 70, 3296–3300 (1973).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  59. Deltcheva, E. et al. CRISPR RNA maturation by trans-encoded small RNA and host factor RNase III. Nature 471, 602–607 (2011).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  60. Maeda, T., Tanaka, Y., Takemoto, N., Hamamoto, N. & Inui, M. RNase III mediated cleavage of the coding region of mraZ mRNA is required for efficient cell division in Corynebacterium glutamicum. Mol. Microbiol. 99, 1149–1166 (2016).

    Article  CAS  PubMed  Google Scholar 

  61. Katzir, I., Cokol, M., Aldridge, B. B. & Alon, U. Prediction of ultra-high-order antibiotic combinations based on pairwise interactions. PLoS Comput. Biol. 15, e1006774 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  62. Hwangbo, H., Patterson, S. C., Dai, A., Plana, D. & Palmer, A. C. Additivity predicts the efficacy of most approved combination therapies for advanced cancer. Nat. Cancer 4, 1693–1704 (2023).

    Article  CAS  PubMed  Google Scholar 

  63. Morreale, F. E. et al. BacPROTACs mediate targeted protein degradation in bacteria. Cell 185, 2338–2353 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  64. Coscolla, M. & Gagneux, S. Does M. tuberculosis genomic diversity explain disease diversity? Drug Discov. Today Dis. Mech. 7, e43–e59 (2010).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  65. Costanzo, M. et al. Environmental robustness of the global yeast genetic interaction network. Science 372, eabf8424 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  66. Typas, A. et al. High-throughput, quantitative analyses of genetic interactions in E. coli. Nat. Methods 5, 781–787 (2008).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  67. Whinn, K. S. et al. Nuclease dead Cas9 is a programmable roadblock for DNA replication. Sci. Rep. 9, 13292 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  68. Beites, T. et al. Plasticity of the Mycobacterium tuberculosis respiratory chain and its impact on tuberculosis drug development. Nat. Commun. 10, 4970 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  69. Murphy, K. C. et al. ORBIT: a new paradigm for genetic engineering of mycobacterial chromosomes. mBio 9, e01467-18 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

  70. Guo, H. et al. Structure of mycobacterial ATP synthase bound to the tuberculosis drug bedaquiline. Nature 589, 143–147 (2021).

    Article  CAS  PubMed  Google Scholar 

  71. Wong, A. I. & Rock, J. M. CRISPR interference (CRISPRi) for targeted gene silencing in mycobacteria. Methods Mol. Biol. 2314, 343–364 (2021).

    Article  CAS  PubMed  Google Scholar 

  72. Dragset, M. S. et al. Genome-wide phenotypic profiling identifies and categorizes genes required for mycobacterial low iron fitness. Sci. Rep. 9, 11394 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  73. DeJesus, M. A. et al. Comprehensive essentiality analysis of the Mycobacterium tuberculosis genome via saturating transposon mutagenesis. mBio 8, e02133-16 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

  74. Deschner, F. et al. Chlorotonils exhibit potent activity against Mycobacterium tuberculosis, while resistance is mediated by MmpR5-MmpL5. Preprint at bioRxiv https://doi.org/10.64898/2025.12.09.693342 (2025).

  75. Zhou, P., Wan, Y. K., Chan, B. K. C., Choi, G. C. G. & Wong, A. S. L. Extensible combinatorial CRISPR screening in mammalian cells. STAR Protoc. 2, 100255 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  76. Hawkins, J. A., Jones, S. K. Jr., Finkelstein, I. J. & Press, W. H. Indel-correcting DNA barcodes for high-throughput sequencing. Proc. Natl Acad. Sci. USA 115, E6217–E6226 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  77. Horlbeck, M. A. et al. Mapping the genetic landscape of human cells. Cell 174, 953–967 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  78. Du, D. et al. Genetic interaction mapping in mammalian cells using CRISPR interference. Nat. Methods 14, 577–580 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  79. Mateyko, N. & de Boer, C. G. Culture wars: empirically determining the best approach for plasmid library amplification. ACS Synth. Biol. 13, 2328–2334 (2024).

    Article  CAS  PubMed  Google Scholar 

  80. Liao, Y., Smyth, G. K. & Shi, W. The Subread aligner: fast, accurate and scalable read mapping by seed-and-vote. Nucleic Acids Res. 41, e108 (2013).

    Article  PubMed  PubMed Central  Google Scholar 

  81. Li, W. et al. MAGeCK enables robust identification of essential genes from genome-scale CRISPR/Cas9 knockout screens. Genome Biol. 15, 554 (2014).

    Article  PubMed  PubMed Central  Google Scholar 

  82. Stan Development Team. Stan Modeling Language Users Guide and Reference Manual v.2.35. https://mc-stan.org/ (2024).

  83. Baryshnikova, A., Costanzo, M., Myers, C. L., Andrews, B. & Boone, C. Genetic interaction networks: toward an understanding of heritability. Annu. Rev. Genomics Hum. Genet. 14, 111–133 (2013).

    Article  CAS  PubMed  Google Scholar 

  84. Wang, T. et al. Pooled CRISPR interference screening enables genome-scale functional genomics study in bacteria with superior performance. Nat. Commun. 9, 2475 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

  85. Hastie, T. J. & Tibshirani, R. J. Generalized Additive Models (Chapman and Hall, 1991).

  86. Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat. Methods 17, 261–272 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  87. Klopfenstein, D. V. et al. GOATOOLS: a Python library for Gene Ontology analyses. Sci. Rep. 8, 10872 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  88. Seshadri, A., Samhita, L., Gaur, R., Malshetty, V. & Varshney, U. Analysis of the fusA2 locus encoding EFG2 in Mycobacterium smegmatis. Tuberculosis 89, 453–464 (2009).

    Article  CAS  PubMed  Google Scholar 

  89. Seshadri, A., Singh, N. S. & Varshney, U. Recycling of the posttermination complexes of Mycobacterium smegmatis and Escherichia coli ribosomes using heterologous factors. J. Mol. Biol. 401, 854–865 (2010).

    Article  CAS  PubMed  Google Scholar 

  90. Mirdita, M. et al. ColabFold: making protein folding accessible to all. Nat. Methods 19, 679–682 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  91. Steinegger, M. & Söding, J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. Biotechnol. 35, 1026–1028 (2017).

    Article  CAS  PubMed  Google Scholar 

  92. Ritz, C., Baty, F., Streibig, J. C. & Gerhard, D. Dose–response analysis using R. PLoS ONE 10, e0146021 (2015).

    Article  PubMed  PubMed Central  Google Scholar 

  93. Zheng, S. et al. SynergyFinder Plus: toward better interpretation and annotation of drug combination screening datasets. Genomics Proteomics Bioinformatics 20, 587–596 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  94. Bliss, C. I. The toxicity of poisons applied jointly. Ann. Appl. Biol. 26, 585–615 (1939).

    Article  CAS  Google Scholar 

  95. Meyer, C. T., Wooten, D. J., Lopez, C. F. & Quaranta, V. Charting the fragmented landscape of drug synergy. Trends Pharmacol. Sci. 41, 266–280 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  96. Ianevski, A., Giri, A. K. & Aittokallio, T. SynergyFinder 3.0: an interactive analysis and consensus interpretation of multi-drug synergies across multiple samples. Nucleic Acids Res. 50, W739–W743 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  97. Brennan-Krohn, T. & Kirby, J. E. When one drug is not enough. Clin. Lab. Med. 39, 345–358 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  98. Teo, G. et al. SAINTexpress: improvements and additional features in significance analysis of interactome software. J. Proteomics 100, 37–43 (2014).

    Article  CAS  PubMed  Google Scholar 

  99. Choi, H. et al. SAINT: probabilistic scoring of affinity purification–mass spectrometry data. Nat. Methods 8, 70–73 (2011).

    Article  CAS  PubMed  Google Scholar 

  100. Buglino, J. A. et al. Diisonitrile lipopeptides mediate resistance to copper starvation in pathogenic mycobacteria. mBio 13, e0251322 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  101. Kresse, I. J. et al. Analysis code for: Comprehensive mapping of essential-gene interactions in M. tuberculosis. Zenodo https://doi.org/10.5281/zenodo.22694611 (2026).

Download references

Acknowledgements

We thank members of the J.M.R. laboratory for comments on the manuscript; J. Chen, L. Perez-Valencia, B. Schultz, M. Kanai, S. Quiñones Garcia, J. Losner, I. Seckler, W. Doyle, M. Xia, A. Carvalloza and the Rockefeller Data Science Platform for co-folding; I. Seckler for protocol development; W. Doyle, R. Froom, F. d’Andrea, M. Kanai, J. Losner, S. Quiñones Garcia, L. Perez-Valencia, K. Eckartt, S. Moraes and K. Bradley for library harvesting; H. Todor and B. Aldridge for discussions; C. Nathan, C. McNamara and J. Sachettini for providing compounds; the Weill Cornell and Rockefeller University Genomics Cores for sequencing; and the Network Biology Collaborative Centre at the Lunenfeld-Tanenbaum Research Institute for mass spectrometry.

Funding

This work was supported by the National Institutes of Health (NIH) NCI Cancer Center Support Grant P30 CA008748 (M.S.G.); the Mathers Foundation (MF-2403-06240 to J.M.R.); the Stavros Niarchos Foundation (SNF) as part of its grant to the SNF Institute for Global Infectious Disease Research at the Rockefeller University (J.M.R.); the Bill and Melinda Gates Foundation (INV-055894 to D.S. and J.M.R.); the Black Family Therapeutic Development Fund (I.J.K., N.Y.S. and J.M.R.); and the Canadian Institutes of Health Research (PJT191893 to J.L.R.). N.Y.S. was supported by a Medical Scientist Training Program grant from the NIH NIGMS (T32GM152349) to the Weill Cornell–Rockefeller–Sloan Kettering Tri-Institutional MD–PhD Program, as well as a predoctoral fellowship from the NIH NIAID (F30AI197420). S.N. was supported in part by an NIH T32 grant (GM132083). J.L.R. was supported by the Canada Research Chairs program. The content of this study is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

Author information

Author notes

  1. These authors contributed equally: Isaac J. Kresse, Natalie Y. Sayegh

Authors and Affiliations

  1. Laboratory of Host-Pathogen Biology, The Rockefeller University, New York, NY, USA

    Isaac J. Kresse, Natalie Y. Sayegh, Vanisha Munsamy-Govender, Allen Na, Samuel Nitz, Barbara Bosch, Sebastian Gonzalez Behn-Eschenburg, Nicholas C. Poulton, Abhijna K. Reddy, Andrew I. Wong, S. Nicolai Tornow, Michael A. DeJesus & Jeremy M. Rock

  2. Immunology Program, Sloan Kettering Institute, Memorial Sloan Kettering Cancer Center, New York, NY, USA

    Yaprak Ozakman & Michael S. Glickman

  3. Department of Microbiology and Immunology, Weill Cornell Medicine, New York, NY, USA

    Joshua Wallach, Rodrigo Aguilera Olvera, Anthony Castro, Heather Kim & Dirk Schnappinger

  4. Tri-Institutional PhD Program in Chemical Biology, Weill Cornell Medicine, The Rockefeller University and Memorial Sloan Kettering Cancer Center, New York, NY, USA

    S. Nicolai Tornow

  5. Molecular Medicine Program, Hospital for Sick Children, Toronto, Ontario, Canada

    Stephanie A. Bueler, Wayne W. H. Law & John L. Rubinstein

  6. Department of Medical Biophysics, University of Toronto, Toronto, Ontario, Canada

    John L. Rubinstein

  7. Department of Biochemistry, University of Toronto, Toronto, Ontario, Canada

    John L. Rubinstein

Authors

  1. Isaac J. Kresse
  2. Natalie Y. Sayegh
  3. Vanisha Munsamy-Govender
  4. Allen Na
  5. Samuel Nitz
  6. Yaprak Ozakman
  7. Barbara Bosch
  8. Sebastian Gonzalez Behn-Eschenburg
  9. Nicholas C. Poulton
  10. Abhijna K. Reddy
  11. Joshua Wallach
  12. Rodrigo Aguilera Olvera
  13. Anthony Castro
  14. Heather Kim
  15. Andrew I. Wong
  16. S. Nicolai Tornow
  17. Stephanie A. Bueler
  18. Wayne W. H. Law
  19. Michael S. Glickman
  20. John L. Rubinstein
  21. Dirk Schnappinger
  22. Michael A. DeJesus
  23. Jeremy M. Rock

Contributions

I.J.K., N.Y.S. and V.M.-G. performed Mtb GI experiments and downstream validation. I.J.K. and N.Y.S. performed Msm GI experiments and downstream validation. S.G.B.-E., B.B., A.I.W., S.N.T., N.Y.S. and I.J.K. developed methods and reagents to quantify GIs. A.N., A.K.R., N.Y.S. and M.A.D. developed computational tools for sgRNA selection and GI visualization. M.A.D. developed statistical models and computational approaches to quantify and analyse GIs. A.N. and I.J.K. analysed GI clusters for enrichment of biological processes. S.N. led the co-folding campaign and generated the associated figures. J.W., R.A.O., A.C., H.K. and D.S. led the Mtb CGI work. Y.O. performed the sensor spot assays. I.J.K., V.M.-G. and N.C.P. performed drug interaction experiments. S.A.B. generated the Ms2764–3×Flag strain and performed the AP–MS experiment. W.W.H.L. generated the ms2764 and qcrB deletion strains and performed the IMV assays. J.L.R. advised on data interpretation, protocol generation and figure preparation. M.S.G., J.L.R., D.S. and J.M.R. supervised experiments. J.M.R. and D.S. conceived the project. I.J.K., N.Y.S., M.A.D. and J.M.R. wrote the manuscript and prepared the figures, with input from all co-authors.

Corresponding authors

Correspondence to Michael A. DeJesus or Jeremy M. Rock.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature thanks David Alland, Clifton Barry, III and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended Data Fig. 1 GI library design.

a, Construction of dual-sgRNA libraries. Single-sgRNA constructs were cloned individually. To assemble the dual-sgRNA library, all single-sgRNA plasmids were pooled and subjected to two parallel digestions: (1) MfeI + XhoI to generate a backbone library, and (2) EcoRI + SalI to excise the sgRNA–barcode fragment and generate an insert library. Ligation of these two products places two sgRNA cassettes in tandem, followed by their corresponding unique barcodes. To quantify sgRNA abundance, the barcode region is amplified using flanking NGS primers. b, Overview of the pilot GI library design. Up to 24 sgRNAs were selected for each of 16 target genes. The resulting dual-sgRNA library was transformed into Msm and serially passaged +/–ATc, with samples collected every 2.5 to 5 generations. The relative abundance of each sgRNA pair was quantified by sequencing the barcode region. c, Density plot of ΔL2FC versus expected L2FC projected at generation 25 for all dual-targeting constructs in the pilot library. Vertical lines mark the expected L2FC window between –10 and –2. d, Heat map of GI scores for all dual-targeting constructs in the pilot library. Each row and column corresponds to a single sgRNA, and each cell represents a dual-targeting construct. sgRNAs are grouped by target gene and, within each gene, are ordered by sgRNA strength, as defined by the L2FC of the sgRNA when paired with a non-targeting sgRNA at 25 generations (weaker sgRNAs at the top/left; stronger sgRNAs at the bottom/right). Gene pairs highlighted in e–h are marked with dashed boxes. e–h, Scatter plots of ΔL2FC versus expected L2FC projected at generation 25 for all dual-sgRNA constructs targeting glgB and glgE (e), ponA1 and ponA2 (f), folA and folP1 (g) and rpoB and macE (h). Vertical lines mark the expected L2FC window between –10 and –2.

Extended Data Fig. 2 GI library quality control.

a, Library metrics. Metrics are shown for the plasmid library after cloning in E. coli and for the corresponding mycobacterial library after transformation into either Msm or Mtb. Mycobacterial libraries were sequenced from the initial (Generation 0, G0) sample of the uninduced (–ATc) condition. Mtb experiment 1 and experiment 2 were derived from the same plasmid library. b,c,g,h,l,m, Heat maps of the Pearson correlation between sgRNA counts across all time points and replicates for the (b,c) Msm, (g,h) Mtb experiment 1 and (l,m) Mtb experiment 2 libraries. For each experiment, the (b,g,l) –ATc samples and (c,h,m) +ATc samples are shown on separate heat maps. d,i,n, Density plots of uncorrected ΔL2FC versus expected L2FC at 25 generations for all dual-targeting constructs in the (d) Msm, (i) Mtb experiment 1 and (n) Mtb experiment 2 libraries. e,j,o, Density plots of corrected ΔL2FC versus expected L2FC at 25 generations for all dual-targeting constructs in the (e) Msm, (j) Mtb experiment 1 and (o) Mtb experiment 2 libraries. Corrected ΔL2FC values are calculated as the residual (observed – GAM-predicted L2FC) for each dual-targeting construct. For more information, see Methods: ‘Quantification of GIs’ and Figs. 1f and 2c and Extended Data Fig. 4c. f,k,p, Density plots of GI score versus combined gene vulnerability for all gene pairs in the (f) Msm, (k) Mtb experiment 1 and (p) Mtb experiment 2 libraries. Combined gene vulnerability is defined as the sum of the vulnerabilities of the two genes in each pair.

Extended Data Fig. 3 Example GIs in Mtb.

a–l, GI scores in Mtb experiment 1 (a–f) and Mtb experiment 2 (g–l) for all genes paired with whiB7 (a,g), with ftsR (b,h), with mtrA (c,i), with ndh (d,j), with ponA1 (e,k) or with dnaE1 (f,l). Arrows mark specific interactions.

Extended Data Fig. 4 Comprehensive GI map of essential genes in Mtb experiment 2.

a, Distribution of observed L2FC at generation 20 for dual-NT, single-targeting, and dual-targeting constructs, as in Fig. 1b. b, Correlation of observed L2FC at generation 20 for single-targeting constructs differing only in sgRNA position. c, Expected versus observed L2FC at generation 25 for all dual-targeting constructs. Black line, x = y; red line, GAM fit. GI scores were computed as the median GAM residual per gene pair. d, Distribution of GI scores across all gene pairs. Shaded regions denote statistically significant (p > 0.5) interactions. e, Heat map of GI scores for all gene pairs, hierarchically clustered by Pearson correlation of GI profiles. Values beyond ±5 are clipped. f, Distribution of pairwise GI profile correlations among all sgRNAs, and among sgRNAs targeting the same gene. g, Number of positive and negative GIs (p > 0.5) per gene. h, GI scores for all genes paired with secA1 (left) or with secF (right). i, GI scores for all genes paired with murA (left) or with cwlM (right). j, Dendrogram of all genes (left), clustered by GI profile correlation. Branches are coloured by nodes enriched (FDR < 0.005) for GO terms by Fisher’s exact test. Representative clusters (right, from top to bottom: fatty acid biosynthesis, arginine biosynthesis, DNA replication, folate metabolism, thiamine metabolism, menaquinone biosynthesis, and protein degradation). k, Scatter plot comparing GI scores in Mtb experiment 1 and Mtb experiment 2, coloured by joint model hit class. See Methods for additional details. l, Distribution of GI scores for gene pairs grouped into 10 equal-width bins by GI profile correlation. Gene pairs separated by five or fewer genes were excluded to minimize potential polar effects. Boxes indicate the interquartile range (IQR), centre lines indicate the median, and whiskers extend to 1.5 × IQR; outliers are not shown.

Extended Data Fig. 5 Comparison of Msm and Mtb hits and enrichment of pathways for GIs.

a–c, GI scores in the Msm GI screen (left) and the Mtb GI screen (right) for all genes paired with crp (a), with fumC (b) or with whiA (c). d,e, Density plots of GI scores in Msm versus Mtb for (d) all homologous gene pairs and (e) homologous gene pairs sharing the same biological process annotation (“category” in Supplementary Table 4, excluding “Unknown”). The total number of gene pairs included in each analysis (n) is indicated in the plot titles. f, Percent of Mtb GIs conserved in Msm. The two gene pair sets defined in d,e were first filtered to retain only Mtb GIs (p > 0.5 in experiment 2). An Mtb GI was considered conserved if the corresponding gene pair in Msm was also called as a significant interaction (p > 0.5) with the same sign (positive or negative). For each set, percent conservation was calculated as the number of conserved GIs divided by the number of Mtb GIs. The analysis was performed twice: once without additional filters (corresponding to the full bar height) and once excluding potential polar effects (gene pairs ≤ 5 genes apart; corresponding to the filled portion). g, Heat map of log2 odds ratio (OR), testing for enrichment or depletion of significant negative GIs (p > 0.5) within pathway pairs in Mtb, as in Fig. 3c. Only results with q < 0.05 are shown. Hollow circles indicate pathway pairs that fail this threshold after removing possible polar effects (gene pairs ≤ 5 genes apart).

Extended Data Fig. 6 Pooled AlphaFold3 analysis of essential genes in Msm and Mtb.

a, Co-folding pool sizes (number of monomers) generated by size-aware bin packing for all gene pairs screened in Msm. All pools have exactly one copy of each monomer (that is, the co-folding screen captures only heterodimers). Six additional pools with >25 monomers not shown. b, Co-folding pool sizes (total residues) generated by size-aware bin packing for all gene pairs screened in Msm. 40 additional pools with fewer than 4,450 residues not shown. c, Co-folding pool sizes (number of monomers) for all gene pairs screened in Mtb, as in a. Five pools with >25 monomers not shown. d, Co-folding pool sizes (total residues) for all gene pairs screened in Mtb, as in b. 36 pools with fewer than 4,450 residues not shown. e, Uncorrected mean ipTM scores for all gene pairs screened in Mtb. Red line is the correction factor, calculated as (0.0044 × \(\sqrt{{l}_{1}\,+\,{l}_{2}}\,\)) − 0.036, where (l1, l2) are the lengths of each protein in the pair. f, Size-corrected (s.c.) mean ipTM scores for all gene pairs screened in Mtb. g,h, Distribution of s.c. ipTMs for co-folded Msm protein pairs with mean s.c. ipTM ≥ 0.2 (g) and all protein pairs (h). Black dots and grey lines represent the mean and s.d. of s.c. ipTMs across all pools.

Extended Data Fig. 7 CGIs and drug–drug interactions in Mtb.

a, Volcano plot showing results from an Mtb CRISPRi-based CGI screen with the KasA and KasB inhibitor thiolactomycin. Knockdown of kasA and pptT sensitizes Mtb to thiolactomycin, whereas knockdown of rpoB does not. FDR was calculated using α-RRA with one-sided tests for enrichment and depletion. b–k, Checkerboard assay heat maps showing Bliss scores for the indicated drug pairs in Mtb, as in Fig. 4g,i,j,l. b, Known synergistic pair: RpoB inhibitor rifampicin (RIF) and EmbA and EmbB inhibitor ethambutol (EMB); c, Known additive pair: ribosome inhibitor linezolid (LZD) and EMB; d, Thiolactomycin (TLM) and LZD; e, TLM and EMB; f, PptT inhibitor 8918 and LZD; g, 8918 and EMB; h, Pks13 inhibitor mCMX410 and LZD; i, EfpA inhibitor BRD-8000.3 and LZD; j, PknA and PknB inhibitor Vertex-49 (V49) and EMB; and k, V49 and ATP synthase inhibitor bedaquiline (BDQ). n = 3 biological replicates.

Extended Data Fig. 8 Highlighted negative GIs in Mtb.

a–e, GI scores for all genes paired with rpoB (a), with atpB (b), with nadA (c), with nadB (d) or with nadD (e). Arrows mark specific interactions.

Extended Data Fig. 9 GIs and validated gene-drug interactions in Mtb.

a–g, Each panel corresponds to a distinct gene: (a) clpX, (b) inhA, (c) dprE1, (d) mmpL3, (e) pks13, (f) glcB, and (g) fusA1. Histograms show the distribution of GI scores for all genes paired with the gene indicated in the plot title. Adjacent plots show dose–response curves (mean ± s.e.m., n = 3 biological replicates) for hypomorphic CRISPRi knockdown (KD) strain(s) and a non-targeting (NT) control strain in the presence of the drug indicated on the x axis. ATc was added five days prior to the assay. Knockdown–drug pairings were selected to test whether negative GIs predict drug sensitization, by combining genetic inhibition of one gene with pharmacological inhibition of its interaction partner. Control drug–KD pairings not predicted to induce sensitization based on GI data are shown on the left and test drug(s) are shown to the right. Drugs include: DnaN inhibitor griselimycin (GRIS); InhA inhibitors isoniazid (INH) and NITD-916; DprE1 inhibitors PBTZ169 and BTZ043; MmpL3 inhibitors AU1235 and SQ109; Pks13 inhibitor mCMX410; GlcB inhibitor JSF1305; FusA1 inhibitor fusidic acid (FUS); and control antibiotics clarithromycin (CLR; 50S ribosome) and streptomycin (STR; 30S ribosome).

Extended Data Fig. 10 Additional evidence supporting roles for previously uncharacterized genes in the ETC.

a, A representative cluster from the GI profile dendrogram in Mtb reveals a coherent module of ETC-associated genes along with several genes of unknown function. Branches are coloured as in Fig. 1l. b, Complementation experiments in a cyt. bd Msm knockout. “Background” indicates the gene targeted by CRISPRi. “Complementation” specifies the contents of each complementation construct, including empty vector controls, CRISPRi-resistant alleles or Mtb homologues of the targeted genes, and in some cases downstream genes to exclude polar effects. Representative image shown from one of two biological replicates. c, GI scores for complex III2IV2 genes, cyt. bd, and strongly correlated genes of unknown function in Mtb. d, Predicted heterodimeric structures generated by AlphaFold-Multimer for selected gene pairs identified as high-confidence physical interactors in the pooled AlphaFold3 analysis. e, Methylene blue oxygen consumption assays of wild-type (WT; bottom) and cyt. bd knockout (Δbd; top) Mtb strains following knockdown of the indicated genes or treatment with 400 nM Q203 (a complex III2IV2 inhibitor). Assays were performed as in Fig. 5l but represent an independent experiment. Imaged four days after adding methylene blue and sealing the vials. Representative image shown from one of three independent experiments. f, CFU concentration in all vials in e at the assay endpoint, determined by serial dilution plating. g, Percent dissolved oxygen after 48 h growth of Mtb cyt. bd knockout strains following knockdown of the indicated genes or treatment with 100 nM Q203, as measured by a sensor spot assay. Each strain was incubated in sealed vials with or without ATc. n = 2 biological replicates, 3 technical replicates each. h, Dose–response curves (mean ± s.e.m., n = 3 biological replicates) for Mtb CRISPRi strains in the presence of linezolid (LZD). ATc was added five days before the assay.

Supplementary information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

[Check for updates. Verify currency and authenticity via CrossMark]

Cite this article

Kresse, I.J., Sayegh, N.Y., Munsamy-Govender, V. et al. Comprehensive mapping of essential-gene interactions in M. tuberculosis. Nature (2026). https://doi.org/10.1038/s41586-026-11212-0

Download citation

  • Received: 10 December 2025

  • Accepted: 24 September 2026

  • Published: 04 November 2026

  • Version of record: 04 November 2026

  • DOI: https://doi.org/10.1038/s41586-026-11212-0