A design approach for bitopic kinase inhibitors

Kinases control many fundamental cellular processes, and their dysregulation underlies many diseases. Small-molecule kinase inhibitors were among the first targeted cancer therapies, and many have been introduced since imatinib was approved in 2001 for chronic myeloid leukaemia (CML) driven by the BCR::ABL1 fusion gene⁴. However, kinase inhibitors often face challenges from resistance mutations and off-target effects that limit their clinical utility.

Linked molecules, built from two affinity ligands connected by a flexible tether, have emerged as a versatile molecular design strategy in recent decades⁵. Although most linked molecules have been developed to engage two different proteins, bitopic inhibitors bind two distinct sites on the same target^1,6. By engaging a larger binding interface, bitopic inhibitors can potentially achieve both higher potency and greater selectivity than conventional single-site inhibitors.

Early bitopic kinase inhibitors linked ATP-competitive ligands to peptide ligands^7,8,9, but poor cell permeability limited their therapeutic potential¹. More recently, linked molecules that connect two drug-like ligands have demonstrated favourable cellular activity, particularly proteolysis-targeting chimeras¹⁰. Our group has designed bitopic inhibitors of mTOR kinase that engage its ATP site directly and its FRB domain through FKBP12¹¹, leading to the development of one of the first bitopic kinase inhibitors to enter clinical trials: RMC-5552^12,13. Recently, we presented a proof-of-concept bitopic inhibitor of ABL1¹⁴), a target that exemplifies the promises and challenges of kinase inhibitors.

The BCR–ABL1 fusion tyrosine kinase drives nearly all CML² and around 30% of adult acute lymphoblastic leukaemia (ALL)¹⁵, which is also associated with other ABL1 and ABL2 fusions^16,17. Although tyrosine kinase inhibitors (TKIs), such as imatinib, have substantially improved patient outcomes in CML¹⁸, two major challenges remain. First, most patients require lifelong therapy, and only approximately 20% of all patients with CML achieve treatment-free remission^19,20. Second, resistance mutations, particularly in the ATP-binding site^2,21, can render TKIs ineffective.

The third-generation TKI ponatinib addresses single-nucleotide resistance mutations of BCR::ABL1, including T315I^2,22. However, this TKI has a US Food and Drug Administration boxed warning for serious cardiovascular events, which are probably caused by off-target inhibition of VEGFR and FGFR kinases³. Ponatinib is ineffective against certain compound BCR::ABL1 mutations such as E255V/T315I, which are more common in advanced disease^23,24,25,26. Conversely, the allosteric inhibitor asciminib improves selectivity by engaging a myristoyl-binding pocket specific to ABL1 and ABL2 kinases in the C-lobe to induce an autoinhibited conformation^27,28. The improved tolerability of asciminib has led to widespread use, and it may improve the rate of response in treatment-naive cases²⁹. However, asciminib is only moderately potent against resistance mutants. It requires fivefold higher clinical dosing for the T315I variant³⁰, is vulnerable to substitutions across the kinase and regulatory domains^31,32,33,34, and is ineffective against mutations such as E255V/T315I^35,36. Although asciminib has not been compared with ponatinib in a head-to-head trial, separate trials suggest that it is less effective than ponatinib as third-line or frontline therapy^37,38. Combined ponatinib and asciminib treatment has shown promise³⁶; however, a recent clinical report suggests that toxicity remains a limiting factor³⁹. A drug that targets the ABL kinases with both the potency of ponatinib and the selectivity of asciminib could improve patient outcomes by enabling deeper molecular responses with fewer adverse effects.

To achieve this goal, we explored a chemical strategy to preserve the potency benefits of dual-site binding⁴⁰ while reducing off-target toxicity. A bitopic inhibitor gains potency and selectivity through avidity: that is, on-target proteins can engage both ligands and benefit from their combined binding interactions, whereas off-target proteins that can only engage one ligand are limited to weaker single-site binding.

Bitopic inhibitors represent a promising strategy to overcome the limitations of traditional single-site inhibitors, but the design principles governing their activity remain poorly understood. Building on our previous proof-of-concept bitopic dasatinib–asciminib compound¹⁴, we sought to develop a generalizable approach to design bitopic inhibitors. Using ABL1 and EGFR as model kinases, we here investigate how active-site ligand choice, linkage vector and linker length influence bitopic inhibitor potency through mechanisms such as inter-ligand cooperativity and linker entropy. We apply these principles to develop PonatiLink-2, a bitopic ABL1 inhibitor with improved activity against severe resistance mutations and reduced off-target toxicity.

We synthesized a panel of bitopic ABL1 inhibitors by linking either dasatinib or ponatinib to asciminib using flexible polyethylene glycol (PEG)-based linkers (Fig. 1a and Extended Data Fig. 1). We divided our library into three compound series to explore several structure–activity relationship parameters: active-site ligand, linkage vector and linker length. Compounds were named according to their series and linker length (for example, PonatiLink-1-PEG24 has a linker consisting of 24 PEG units).

To evaluate the effect of the active-site ligand, we synthesized both dasatinib-based (DasatiLink) and ponatinib-based (PonatiLink) series. We first tested the potency of the parent compounds dasatinib, ponatinib and asciminib in K562 cells derived from a patient with CML, which express wild-type BCR::ABL1, and in K562 cells that were base-edited to endogenously express a mutant (BCR::ABL1^T315I). Dasatinib was highly potent against cells expressing wild-type BCR::ABL1 but practically inactive against the T315I mutant. Ponatinib was highly potent against both cell lines, and asciminib was moderately potent against the wild-type and less potent against the mutant (Extended Data Fig. 1c).

To explore linkage vector effects, we took advantage of the structure of ponatinib, which spans the width of the ABL1 ATP-binding pocket and enables linker attachment on either side of the molecule⁴¹. Our PonatiLink-1 series is linked through the piperazine group, whereas the PonatiLink-2 series is linked through the imidazopyridazine heterocycle. To facilitate synthesis, we replaced the imidazopyridazine moiety with an imidazopyridine moiety, a modification reported to minimally affect the potency of ponatinib⁴². All three series included analogues with various linker lengths (Extended Data Fig. 1d).

We evaluated the effects of active-site ligand, linkage vector and linker length by testing compounds against purified full-length wild-type ABL1 and ABL1(T315I) and against K562 cells expressing wild-type BCR::ABL1 or BCR::ABL1^T315I. We compared bitopics to their parent inhibitors and to 1:1 mixtures of ponatinib and asciminib (pona+asc) or of dasatinib and asciminib (dasa+asc; for example, 10 nM pona+asc indicates 10 nM ponatinib and 10 nM asciminib). All three series were active in cells, with half-maximal inhibitory concentration (IC₅₀) values of <10 nM (Extended Data Fig. 2). Longer linkers generally increased biochemical potency, whereas the longest PonatiLink compounds were less potent in cells. This result suggests that there is a trade-off between biochemical affinity and cell permeability. Therefore, we selected long but not maximal linker lengths for each series as lead compounds for subsequent study.

We next compared the lead compounds of each series to identify an overall lead (Fig. 1b,c). Hereafter, we refer to series lead compounds by their series name: DasatiLink-1-PEG12 as DasatiLink-1; PonatiLink-1-PEG24 as PonatiLink-1; and PonatiLink-2-PEG17 as PonatiLink-2. DasatiLink-1, although less potent than dasatinib against wild-type ABL1 kinase and K562 cells, was more potent than dasatinib against the T315I mutant. PonatiLink-2 (Fig. 1d) was the most potent series lead in K562 cells expressing BCR::ABL1^T315I and in purified ABL1(T315I), for which it also exceeded the potency of ponatinib.

To assess the contribution of avidity to the affinity of PonatiLink-2 for ABL1, we performed KINOMEscan binding assays. In these assays, an inhibitor of interest competes with bead-immobilized promiscuous ATP-competitive ligands for binding to ABL1 (ref. ⁴³). Binding of the competitor displaces ABL1 from the beads, which enables dose–response measurements to estimate the apparent affinity (dissociation constant, K_d) of the competitor. A 100-fold excess of asciminib did not affect the apparent affinity of ponatinib (0.45 nM) at the ATP site; however, it decreased the apparent affinity of PonatiLink-2 by 20-fold (0.67 nM to 13.9 nM) (Extended Data Fig. 2g). These results suggest that avidity contributes substantially to the potency of PonatiLink-2, with both ligands engaging a single ABL1 protomer.

To investigate the mechanism behind differences in potency between dasatinib-based and ponatinib-based bitopics (DasatiLink-1 was less potent than dasatinib, whereas PonatiLink-2 was more potent than ponatinib), we performed bioluminescence resonance energy transfer (BRET) assays based on NanoLuc (NanoBRET) to measure how active-site ligands affect binding of an asciminib-based fluorescent tracer (asc-tracer)³⁴ to full-length ABL1 fused to luciferase in live HEK293T cells⁴⁴ (Extended Data Figs. 3 and 4). Because orthosteric inhibitors bind a distinct site, they do not directly compete with asc-tracer binding. Instead, changes in the apparent dissociation constant (K_d,apparent) of asc-tracer in the presence of excess orthosteric inhibitor indicated ligand binding cooperativity: that is, positive cooperativity decreases K_d,apparent, whereas negative cooperativity increases K_d,apparent, and no cooperativity leaves K_d,apparent unchanged. Ligands with more positive cooperativity might be expected to be more effective in a bitopic inhibitor.

We compared the effect of clinical active-site ABL1 inhibitors on the K_d,apparent of asc-tracer to a pair of control compounds previously reported as conformation-selective binders of ABL1: DAS-CHO-II and DAS-DFGO-II. DAS-CHO-II binds the C-helix-out (CHO) conformation of the kinase domain, whereas DAS-DFGO-II binds the DFG-out (DFGO) conformation^45,46. Asciminib stabilizes the SH3–SH2-domain closed state of ABL1, which induces the CHO conformation⁴⁶. DAS-CHO-II showed positive cooperativity with asc-tracer as expected (0.86-fold change in K_d,apparent), whereas DAS-DFGO-II and imatinib demonstrated negative cooperativity (2.7-fold and 3-fold, respectively, change in K_d,apparent).

Dasatinib, which binds the active DFG-in state of the kinase domain, had negative cooperativity similar to DAS-DFGO-II (2.5-fold). Conversely, ponatinib, which is reported to bind the DFGO conformation⁴¹, displayed less negative cooperativity (1.7-fold; Fig. 1e). To validate these findings, we repeated our experiments using the opposite ligand as the fluorescent tracer. We measured the K_d,apparent of a dasatinib-based tracer (das-tracer) and a ponatinib-based tracer (pon-tracer-2), both with the same linkage vector as PonatiLink-2, in the presence of excess asciminib. We observed the same pattern: pon-tracer-2 showed less negative cooperativity (2.1-fold) with asciminib than das-tracer (4.7-fold; Extended Data Fig. 3c). These findings indicate that differences along a spectrum of ligand binding cooperativity may be consequential for bitopic inhibitor binding. That is, the reduced negative cooperativity of ponatinib with asciminib may contribute to the enhanced potency of ponatinib-based bitopics.

To assess the broader applicability of our findings, we applied a similar approach to design bitopic inhibitors of EGFR, a clinically important kinase target in non-small-cell lung cancer and other solid tumours⁴⁷ (Fig. 2a and Extended Data Fig. 6). We synthesized a series of compounds based on two EGFR inhibitors: osimertinib, a covalent orthosteric inhibitor in which the acrylamide moiety reacts with Cys797 (ref. ⁴⁸); and RO7304898, a reversible allosteric inhibitor^49,50. The binding sites of these inhibitors are physically adjacent, and osimertinib and allosteric inhibitors can bind EGFR simultaneously⁵¹. However, directly linking two EGFR inhibitors through the pocket interior while maintaining favourable geometry for binding has proven challenging^52,53. We pursued an alternative strategy, whereby longer, solvent-exposed linkers were used to traverse the kinase exterior while preserving ligand-binding geometry.

We prepared OsimertiLink-1 compounds with linkers of 4–25 PEG units and measured their covalent labelling of purified EGFR by intact-protein mass spectrometry. The compounds were compared with osimertinib with its linker attachment point acetylated (Ac-osimertinib) as a control for any change in reactivity caused by linker attachment. OsimertiLink-1-PEG4, which has a linker too short to enable simultaneous binding to the same protomer, labelled EGFR more slowly than Ac-osimertinib. By contrast, OsimertiLink-1-PEG11, and all the other compounds with linkers sufficiently long for simultaneous binding, labelled EGFR at similar rates, faster than Ac-osimertinib (Fig. 2b,c). In this instance, covalent labelling was not dominated by the intrinsic reactivity of the electrophilic component, which remained constant throughout the OsimertiLink-1 series. These results suggest that dual-site engagement on a single protomer is important for bitopic inhibitor activity.

Although OsimertiLink-1 compounds successfully labelled EGFR, we observed that Ac-osimertinib labelled more slowly than osimertinib itself (Extended Data Fig. 6b). We reasoned that optimizing the linkage chemistry by restoring amine rather than amide functionality at the linker attachment point to mimic the tertiary alkyl amine of osimertinib might produce more active compounds. Using the OsimertiLink-1 series as a guide for optimal linker length, we designed and synthesized OsimertiLink-2-PEG10 (Fig. 2d and Extended Data Fig. 6c), hereafter referred to as OsimertiLink-2.

OsimertiLink-2 labelled EGFR faster than any OsimertiLink-1 compound (Fig. 2b). Moreover, OsimertiLink-2 was active in cells. That is, it inhibited the viability of H1975 cells, which are driven by EGFR^T790M/L858R, less potently but with greater maximal inhibition than osimertinib (IC₅₀ values of 133 nM and 10 nM, respectively; Fig. 2e). OsimertiLink-2 exhibited minimal off-target toxicity against A549 cells, which are also derived from lung adenocarcinoma but driven by the KRAS^G12S mutation, which should render them insensitive to EGFR inhibition. A piperazine derivative of RO7304898 (RO-amine) was used for comparison instead of RO7304898 itself, as it was readily available as an intermediate in the OsimertiLink-2 synthesis process. To assess on-target activity against mutant EGFR, we treated H1975 cells with OsimertiLink-2 or parent compounds and measured the phosphorylation of EGFR (p-EGFR(Y845)) and its downstream effectors⁴⁷ AKT (p-AKT(S473)) and ERK1 and ERK2 (p-ERK(T202/Y204)) by western blotting (Fig. 2f). All treatments induced expression of BIM, an indicator of apoptosis, but OsimertiLink-2 less so than osimertinib. OsimertiLink-2 inhibited EGFR pathway phosphorylation with a potency similar to osimertinib. This result suggests that there is comparable target engagement despite differences in growth inhibition potency.

Combination treatment with ponatinib and asciminib has been shown to restore efficacy against some BCR::ABL1 mutants that are resistant to ponatinib or asciminib separately³⁶. We investigated whether PonatiLink-2 was similarly effective by comparing it to ponatinib and asciminib, alone and in combination, in their ability to inhibit growth of a pooled saturation mutagenesis library of K562 cells lentivirally overexpressing single-amino-acid variants across the BCR–ABL1 SH3, SH2 and kinase domains (Fig. 3a,b and Extended Data Fig. 7a). Overall, 94% of all possible variants were detected in the cell pool before compound treatment. Neither ponatinib nor asciminib was able to inhibit outgrowth of the library at clinically realistic concentrations, whereas combined ponatinib and asciminib effectively suppressed growth. PonatiLink-2 suppressed growth of the library at 1 μM, an exposure that seems achievable in vivo (see below and Fig. 5).

To validate these findings, we tested K562 cell lines expressing individual clinically relevant resistance mutants: the active-site mutants E255V and T315I, which are resistant to imatinib but sensitive to ponatinib; the ponatinib-resistant active-site mutants E255V/T315I and T315M²⁴; and the asciminib-resistant allosteric-site mutant V468F⁵⁴ (Fig. 3c). We also tested compound mutants of V468F with ponatinib-resistant active-site mutants recently observed clinically⁵⁵, thereby representing potential resistance to pona+asc.

We quantified inhibitor activity as the ratio of the IC₅₀ value against a mutant cell line to the IC₅₀ value against cells expressing wild-type BCR::ABL1 (Fig. 3d and Extended Data Fig. 7b–d). Such IC₅₀ changes can be interpreted in the context of known clinical resistance. For example, ponatinib is clinically effective against BCR–ABL1(T315I) but not BCR–ABL1(T315M), which suggests that mutants with IC₅₀ shifts similar to T315M are also likely to be resistant²⁴. Although clinical resistance to a new inhibitor also depends on its achievable exposure and other pharmacological characteristics, fold changes provide a starting point for predicting efficacy.

Ponatinib maintained its potency, with IC₅₀ changes of less than fivefold against all mutants tested except those known to be clinically resistant (E255V/T315I and T315M). Asciminib was active against wild-type and E255V, but its potency was greatly reduced in all other mutants tested. Consistent with previous reports³⁶, pona+asc was at least as effective as ponatinib or asciminib against every mutant tested and modestly more effective against E255V/T315I and T315M than ponatinib alone. E255V/T315I/V468F and T315M/V468F, which should theoretically block both ponatinib and asciminib binding, were indeed highly resistant to ponatinib, asciminib and pona+asc.

PonatiLink-2 was highly effective against active-site resistance mutants, even those resistant to ponatinib. E255V/T315I and T315M, which are clinically resistant to ponatinib and increased its IC₅₀ value by 79-fold and 33-fold, respectively, increased the IC₅₀ value of PonatiLink-2 by only 5.8-fold and 4.7-fold, respectively. V468F, which is highly resistant to asciminib³⁵, increased the IC₅₀ of PonatiLink-2 by only 4.4-fold, comparable to the effect of E255V on pona+asc. By contrast, compound mutants affecting both binding sites, such as T315M/V468F, were highly resistant, which indicates that PonatiLink-2 acts on target by inhibiting ABL1. Lentiviral expression did not appreciably increase the total level of BCR–ABL1 in representative mutant cell lines (Extended Data Fig. 7e).

To assess the potency of PonatiLink-2 against BCR–ABL1 signalling, we treated K562 cells lentivirally overexpressing wild-type BCR::ABL1 or BCR::ABL1^E255V/T315I with PonatiLink-2 or parent compounds and measured the phosphorylation of ABL1 (p-ABL1(Y245)) and its known targets STAT5 (p-STAT5(Y694)) and CRKL (p-CRKL(Y207))^21,56 by western blotting (Fig. 3e). In cells expressing wild-type BCR::ABL1, PonatiLink-2 inhibited phosphorylation with a potency similar to pona+asc and slightly exceeding that of ponatinib. In cells expressing BCR::ABL1^E255V/T315I, PonatiLink-2 exceeded the potency of ponatinib and pona+asc.

We investigated the kinase selectivity of PonatiLink-2 (Fig. 4a) and pona+asc across a SelectScreen panel of 386 purified human kinases (Fig. 4b). We tested each treatment at its previously determined IC₉₀ for ABL1(T315I). Consistent with previous reports²², pona+asc had potent off-target activity, whereby the combination inhibited 39 kinases, 10% of the panel, by at least 65% (S_65% = 0.10)⁴³. These included kinases in the FLT, VEGFR, PDGFR and FGFR families, which are suspected to be responsible for the vascular toxicity of ponatinib³.

PonatiLink-2 was substantially more selective: it did not inhibit any off-target kinase by more than 38% (EPHA8), with the exception of 80% inhibition of ABL2, which is also inhibited by asciminib^32,57. ABL1 and ABL2 constituted 0.5% of the kinases tested (S_65% = 0.005).

To assess the haematopoietic toxicity of PonatiLink-2, we treated primary human peripheral blood mononuclear cells, obtained from a healthy individual, with PonatiLink-2 or its parent clinical inhibitors for 14 days (Fig. 4c and Extended Data Fig. 8). PonatiLink-2 and asciminib demonstrated minimal toxicity at up to 10 μM, which was in contrast to the marked toxicity of ponatinib and pona+asc. This finding indicates that PonatiLink-2 has improved tolerability in primary human cells.

To more broadly investigate off-target toxicity across cell types, we compared PonatiLink-2 with ponatinib and asciminib in the PRISM cancer cell line screen, which measures the viability of pooled DNA-barcoded cell lines after compound treatment⁵⁸ (Fig. 4d,e). We tested each compound at identical doses from 0.9 nM to 2 μM. A total of 865 cell lines passed quality checks, including ten BCR::ABL1⁺ lines originating from CML, one BCR::ABL1⁺ line from ALL and one from T-lymphoblastic leukaemia (TLL) driven by NUP214::ABL1, the second-most-common oncogenic ABL1 fusion after BCR::ABL1 (ref. ⁵⁹).

All three compounds were selective for cell lines driven by ABL1 fusions, but asciminib and PonatiLink-2 were more selective than ponatinib. Asciminib did not inhibit any of the non-ABL1-driven cell lines more than the least sensitive ABL1-driven line. PonatiLink-2 inhibited only two (EOL1, an AML line driven by PDGFRα that was also highly sensitive to ponatinib, and NP3, a glioblastoma line), whereas ponatinib inhibited 28. Ponatinib also exhibited greater toxicity across the remaining cell lines, with an average area under the curve (AUC) of 0.76 compared with 0.88 for asciminib and 0.90 for PonatiLink-2. At the highest tested dose of 2 μM, median viability across all cell lines was 26% for ponatinib compared with 101% for asciminib and 97% for PonatiLink-2.

To model clinically relevant off-target cardiovascular toxicity, we used human umbilical vein endothelial cells (HUVECs), as the cardiovascular side effects of ponatinib are thought to be mediated by vascular toxicity rather than direct effects on cardiomyocytes⁶⁰. We treated HUVECs with PonatiLink-2 or its parent clinical inhibitors, comparing the ratio of the IC₅₀ value of each compound against HUVECs to its IC₅₀ value against K562 cells expressing wild-type BCR::ABL1 (Fig. 4f). Ponatinib and pona+asc were 120–140-fold less potent against HUVECs than against K562 wild-type BCR::ABL1. Asciminib was more selective, with a 3,400-fold IC₅₀ difference, and PonatiLink-2 was even more so, with a 9,100-fold window, which suggests that it has reduced off-target vascular toxicity.

We compared PonatiLink-2 with established therapies in mouse models of wild-type BCR::ABL1 or highly resistant mutant CML (Fig. 5 and Extended Data Fig. 9). To model wild-type BCR::ABL1 CML, we used mice with xenografts of K562 cells. We compared PonatiLink-2 with dasatinib, a common frontline treatment option associated with increased rates of deep molecular responses likely to enable treatment-free remission⁶¹. Dasatinib was administered in a regimen (5 mg kg^–1 orally) daily (Monday–Friday) reported to ‘closely mimic the pharmacokinetics of 100 mg oral dose in human’⁶². We compared this regimen to the same dasatinib schedule plus PonatiLink-2 (100 mg kg^–1 intraperitoneally) twice a week on consecutive days. Treatment was withdrawn after 3 weeks, and mice were observed for three additional months. Dasatinib significantly prolonged time to tumour size end point compared with vehicle control (P = 0.018) but induced durable off-treatment responses in only 4 out of 8 mice. By contrast, the dasatinib and PonatiLink-2 combination significantly prolonged time to tumour size end point (P < 0.001) and induced durable off-treatment responses in 8 out of 8 mice (Fig. 5a,b).

To model ponatinib-resistant CML, we used mice with xenografts of K562 cells engineered to express BCR::ABL1^E255V/T315I. We used a high but tolerated dose of ponatinib (30 mg kg^–1)²², using ponatinib alone because the combination with asciminib is not clinically established and may be limited by toxicity³⁹. We compared it to PonatiLink-2, dosed daily at the same level as in the wild-type BCR::ABL1 model. Ponatinib delayed tumour growth slightly but had no significant effect on the time to tumour size end point. By contrast, PonatiLink-2 had a highly significant effect (P = 0.004), with no mice reaching the end point by the end of the 6-week dosing period (Fig. 5c,d).

A single tumour grew to appreciable but sub-end-point size under treatment with PonatiLink-2. A cell line cultured from the resistant tumour (K562 PL2R) displayed substantial additional resistance to asciminib (77-fold IC₅₀ increase) and PonatiLink-2 (23-fold), a slight resistance to ponatinib (3.2-fold) and minimal resistance to RMC-5552 (1.7-fold), a non-ABL1-targeting bitopic inhibitor that we included to test for nonspecific large-molecule-resistance mechanisms (Extended Data Fig. 10). Sequencing of lentivirus-integrated ABL1 revealed, in addition to the parental E255V and T315I mutations, two new spontaneous mutations in the resistant tumour and cell line: G463D and E505K. Both of these mutations are located in the binding site of asciminib and have been reported to confer resistance to asciminib or its predecessor compound GNF-2 (refs. ^54,57,63,64). The mutations appeared at 50% frequency in the bulk tumour sample, but at almost 100% in the isolated cell line, which indicated that both were present in a single clone. This result suggests that although paths to PonatiLink-2 resistance exist through compound on-target mutations, as observed here from an already highly resistant background, they may rely on mutations that are individually vulnerable to the bitopic inhibitor.

We measured concentrations of the compounds in plasma after 2 weeks of dosing in the ponatinib-resistant model, both immediately before dosing (C_min) and at the approximate time of peak concentration (C_max). Ponatinib reached a C_max of 731 nM but decreased to a C_min of only 42 nM between doses, lower than its IC₅₀ of 83 nM against the K562 BCR::ABL1^E255V/T315I cell line (Extended Data Fig. 7d). By contrast, PonatiLink-2 reached a C_max of 8,820 nM and decreased to a C_min of 4,830 nM, substantially higher than its IC₅₀ of 19 nM against the same. This higher exposure was not accompanied by any apparent gross toxicity, and mice treated with PonatiLink-2 gained body weight over the course of the study (Fig. 5e and Extended Data Fig. 9e).

We based PonatiLink-2 dosing on a pharmacokinetic study, in which mice treated daily for 2 weeks (100 mg kg^–1 intraperitoneally) demonstrated some skin scruffiness but no other adverse effects, and in which high concentrations of PonatiLink-2 persisted for days, with 890 nM detected in plasma a full week after the end of dosing (Extended Data Fig. 9f). In line with other large multivalent inhibitors⁶⁵, PonatiLink-2 is anticipated to exhibit limited oral bioavailability. Notably, the bitopic inhibitor RMC-5552 is administered parenterally¹³, which demonstrates that effective exposure can be achieved through this route. We therefore pursued intraperitoneal administration for PonatiLink-2. These pharmacokinetic data demonstrate that PonatiLink-2 can achieve more than 10-fold higher sustained systemic exposure than ponatinib, which suggests a markedly expanded therapeutic window.

In this work, we investigated the design principles that govern bitopic inhibitors, using ABL1 and EGFR kinases as model systems. By varying active-site ligand, linkage vector and linker length, we identified how these parameters influence potency through ligand-binding cooperativity and linker entropy. We applied these insights to develop PonatiLink-2, a bitopic ABL1 inhibitor with improved selectivity and enhanced potency against resistance mutations.

Our structure–activity relationship analyses revealed design considerations specific to bitopic inhibitors that extend those of traditional medicinal chemistry. Whereas conventional drug design often seeks a lock-and-key fit in a single binding pocket, bitopic inhibitors benefit from linker conformational flexibility. That is, longer linkers enhance potency even beyond the minimum length required to span both binding sites. Our simulations indicated that longer linkers were more mobile in the bound state, which suggests that there is a reduced entropic penalty of binding. Moreover, PonatiLink-2 compounds consistently outperformed PonatiLink-1 compounds of similar length, which may be because their more direct attachment vector allows greater linker flexibility.

Studies of covalent bitopic inhibitors of EGFR, an unrelated kinase target, offered a complementary system to investigate these principles of linker length and to test the generality of the approach. We observed a marked threshold in labelling kinetics at a length sufficient to span the two binding sites and enable simultaneous binding. These findings are consistent with a classic study⁶⁶ showing a similar threshold linker length effect on affinity, and with our observations that the affinity of PonatiLink-2 depends on dual-site avidity. Notably, an optimized analogue, OsimertiLink-2, displayed activity against mutant-EGFR-driven H1975 cancer cells, with potentially reduced toxicity against EGFR-independent A549 cells compared with osimertinib.

Finally, our results suggest that ligand cooperativity, and specifically its magnitude, is another relevant bitopic parameter. Ponatinib displayed less negative cooperativity with asciminib than dasatinib, which may account for the better performance of ponatinib-based over dasatinib-based bitopics relative to their parent compounds. These insights, although not comprehensive, should provide guidance for medicinal chemists developing bitopic inhibitors of various targets and motivate future mechanistic investigations.

Our studies of BCR::ABL1-driven leukaemia demonstrate the therapeutic potential of the bitopic approach. The improved selectivity of PonatiLink-2 enabled substantially higher systemic exposure than ponatinib without apparent toxicity. Our results also demonstrated that PonatiLink-2 leads to sustained exposure, which suggests that it may be suitable for intermittent administration, similar to the clinical bitopic mTOR inhibitor RMC-5552 (ref. ¹³). These favourable properties provide support for two distinct treatment strategies: in a wild-type BCR::ABL1 CML model, intermittent PonatiLink-2 dosing combined with daily dasatinib increased the rate of durable off-treatment response from 50% with dasatinib alone to 100%. By contrast, in a compound-mutant CML model, daily PonatiLink-2 dosing effectively treated tumours that were highly resistant to ponatinib. Potent, selective bitopic inhibitors may find clinical application as monotherapies for resistant disease or as front-line agents. They may also serve as well-tolerated adjuncts for standard therapies to suppress resistance and improve the odds of treatment-free remission.

These findings establish a generalizable design approach for bitopic inhibitors. In this study, we built bitopics from highly optimized parent compounds. However, the potency of PonatiLink-2 against resistance mutants that substantially weaken the binding of either of its ligands suggests that this approach also holds promise for targets with only low-affinity binders available. We hope that this work will facilitate further exploration of multivalent molecules and will prove fruitful for designing potent, selective therapeutics for ABL1 and other targets.

Inhibitors for in vitro and cell-based assays were obtained as follows: ponatinib (MedChemExpress (MCE), HY-12047); asciminib (MCE, HY-104010); imatinib (MCE, HY-15463); dasatinib (MCE, HY-10181); osimertinib (MCE, HY-15772); and RMC-5552 (MCE, HY-132168). DasatiLink-1 and asc-tracer were obtained from the same batches as previously described^14,34. DAS-DFGO-II and DAS-CHO-II were provided by M. Soellner. Das-tracer⁴⁴ was provided by Promega. Synthesis and characterization of all other compounds are described in Supplementary Note 1. All compounds were stored at −20 °C as solids or stock solutions in DMSO.

Cell culture

All cells were cultured in humidified incubators at 5% CO₂ and 37 °C on tissue-culture-treated plasticware. Cells were counted using a Countess II FL automated cell counter (Thermo Fisher Scientific) or manually using a haemocytometer. Cell lines were used directly from ATCC without further authentication, except that HEK293T cells were authenticated by STR testing. HUVECs were used directly from ATCC without mycoplasma testing. HEK293T cells tested negative for mycoplasma before use. K562 cells periodically tested negative for mycoplasma, including K562 PL2R tumour-derived cells. Other cell lines were not tested for mycoplasma during this work.

All K562 cell lines were cultured in RPMI 1640 medium with l-glutamine (Gibco 11875119) and 10% heat-inactivated fetal bovine serum (FBS, Axenia Biologix or Atlas Biologicals) and were passaged every 2–3 days generally by simple dilution into fresh complete medium or, if the medium seemed acidified, by centrifuging (300g, 3 min), aspirating medium and resuspending in fresh medium. Cell density was maintained between 5 × 10⁴ and 1 × 10⁶ cells per ml.

K562 wild-type cells were obtained from the American Type Culture Collection (ATCC, CCL-243). K562 stable cell lines were generated as follows: mutations in the pUltra BCR::ABL1 vector were generated using KLD Enzyme Mix (NEB M0554S) with KOD Xtreme Hot Start DNA Polymerase (Sigma-Aldrich 71975-M). Mutations were confirmed by Sanger sequencing (Azenta) and whole-plasmid sequencing (Plasmidsaurus). The pUltra BCR::ABL1 plasmids used in this work are available from Addgene as 210432 (wild type), 210433 (E255V mutation), 210434 (T315I mutation), 210435 (V468F mutation), 210436 (E255V/T315I mutation), 210437 (T315M mutation), 210438 (E255V/V468F mutation), 210439 (T315I/V468F mutation), 210440 (E255V/T315I/V468F mutation) and 210441 (T315M/V468F mutation). Lentivirus was generated as previously described^67,68. In brief, 1 × 10⁶ HEK293T cells were transfected with 5 μg pUltra BCR::ABL1 and 1 μg of each lentiviral helper plasmid (TAT, GAG-POL, VSV-G and REV) using calcium phosphate supplemented with DEAE dextran in a 6-well plate, and medium was changed to RPMI 1640 after 1 day. Next, 3 × 10⁵ K562 cells were seeded in a 6-well plate and infected with lentivirus at low multiplicity of infection (<0.3) so that most cells would have a single lentiviral integration event. Five days after infection, a pure population of K562 pUltra EGFP-only control and pUltra wild-type BCR::ABL1 cells was selected by sorting on eGFP by the Penn State Flow Cytometry Core Facility on a Beckman Coulter MoFlo Astrios EQ cell sorter. K562 pUltra BCR::ABL1^V468F cells were selected using 100 nM asciminib until >95% of the cells were eGFP-positive. All other K562 pUltra BCR::ABL1 mutants were selected by 1 μM imatinib until >95% of the cells were eGFP-positive. eGFP signal and density were determined using a BD Accuri C6 Plus with standard filter settings. All mutant lines were allowed to recover for at least 3 days after drug selection before use in viability assays.

K562 BCR::ABL1^T315I base-edited BE-T315I cells were generated as previously described⁶⁹. In brief, 5 × 10⁶ K562 wild-type cells were electroporated in a Lonza 4D-Nucleofector X unit with 5 μg Lenti sgABL1_T315I (sgRNA gTCACTGAGTTCATGACCTAC, cloned into Addgene, plasmid 104991) and 5 μg pSI-625 TargetACEmax (Addgene, plasmid 139105) in Chicabuffer 2 M. After 3 days of recovery in complete medium, these electroporated cells were selected in 1 μM imatinib for 2 weeks. gDNA was purified using a Monarch Genomic DNA Purification kit (NEB, T3010L). Exon 6 of ABL1 was amplified using KOD Hot Start polymerase (Sigma-Aldrich 71842) and primers (forward: TCTCAGGATGCAGGTGCTTG; reverse: TGAGTGGCCATGTACAGCAG) designed by Primer-BLAST⁷⁰. Sanger sequencing was performed by the Penn State Genomics Core Facility using the forward amplification primer. Sanger sequencing confirmed the T315I mutation (from ACT to ATT) at a proportion of approximately 31% and a silent mutation at position 314 (from ATC to ATT) at a proportion of approximately 55%, as estimated by EditR (v1.0.10)⁷¹ (Supplementary Data 1). We note that parental K562 cells have genomic amplification of the BCR::ABL1 allele^72,73, so these proportions probably represent multiple but incomplete edits of the population of BCR::ABL1 alleles.

HUVECs were obtained from the ATCC (CRL-1730) and cultured in F-12K medium (ATCC 30-2004) with 10% FBS (Gibco 10082147), 0.1 mg ml^–1 heparin (Sigma-Aldrich H3393) and 0.03 mg ml^–1 endothelial cell growth supplement (Corning 354006). Plasticware was prepared by coating with gelatin (Sigma-Aldrich G9391) as follows: a 2% gelatin solution was added to plates (for example, 0.5 ml for a 6-well plate well) and plates were dried ≥30 min at 37 °C. HUVECs were passaged after reaching 70–80% confluence by washing with PBS, detaching with 0.25% trypsin and washing gently with complete medium before diluting to split 2–3-fold.

HEK293T cells were obtained from the ATCC (CRL-3216) and cultured in DMEM with glucose, glutamine and pyruvate (Corning 10-031-CV), 10% FBS (Avantor 97068-085) and antibiotic–antimycotic (Corning 30-004-CI). Cells were passaged every 2–3 days by washing with PBS (Corning 21-040-CV), detaching with 0.25% trypsin (Corning 25-053-CI) and washing with complete medium before splitting to maintain confluence between 10 and 90%.

A549 cells were obtained from the ATCC (CCL-185) and cultured in DMEM, glucose, glutamine and pyruvate (Gibco 11995065) with 10% FBS (Atlas Biologicals) and 1% penicillin–streptomycin (Gibco 15140-122), and handled as for HEK293T cells.

H1975 cells were obtained from the ATCC (CRL-5908) and cultured in RPMI 1640 with l-glutamine (Gibco 11875093), 10% FBS (Atlas Biologicals) and penicillin–streptomycin (Gibco 15140122), and handled as for HEK293T cells.

In-cell NanoBRET affinity assays

NanoLuc–ABL1 (full-length ABL1, NLuc–ABL1, Promega NV1011) was transfected into HEK293T cells using FuGENE HD (Promega E2311). Cells were plated into 384-well plates (Corning) at a density of 2 × 10⁵ cells per ml in DMEM (Corning 10-031-CV) without FBS and allowed to recover for 24 h. In tracer EC₅₀ experiments, serially diluted tracer compound was added to cells. In unlabelled compound affinity experiments, serially diluted compound was added to cells in addition to tracer compound at the tracer EC₅₀ value previously measured for NLuc–ABL1. In tracer competition-binding experiments, the serially diluted tracer compound was added to cells in addition to unlabelled compound dosed at 100× the largest observed EC₅₀ value across independent affinity experiments with the dasatinib-based tracer. Following addition of tracer compound and/or unlabelled compound in all experiments, cells were equilibrated at 37 °C and 5% CO₂ for 2 h. BRET signals were then recorded using NanoBRET NanoGlo Substrate and Extracellular NanoLuc Inhibitor (Promega, 2160) on a Synergy Neo2 plate reader (Agilent BioTek). The average BRET ratio of the no-tracer background control was subtracted from the BRET ratio of the associated condition, and data were fit with the doseplotr R package as for other dose–response data. Apparent compound affinities with das-tracer were measured with tracer at its EC₅₀ concentration. Asciminib, DAS-DFGO-II and DAS-CHO-II affinity measurements were obtained over two independent experiments; the largest observed EC₅₀ value between independent experiments was used for 100× EC₅₀ calculations.

MD simulations and analysis

Initial structures for the ABL1 kinase domain (residues 255–531) in complex with ligands (asciminib, ponatinib and nilotinib) were taken from the bound forms (PDB identifiers 5MO4 and 3OXZ). Ponatinib (PDB identifier 3OXZ) was transplanted into the asciminib–nilotinib-bound structure (PDB identifier 5MO4) in place of nilotinib after backbone alignment. AlphaFold3 (ref. ⁷⁴) was used to generate a complete template of ABL1, which was then used with MODELLER⁷⁵ to fill in unresolved regions of the experimentally determined structures. Linker placement and construction were performed in MODELLER using glycine residues as a proxy for the PEG unit, with anchor points defined at the attachment points of the PEG linker. The PEG linker coordinates were mapped out onto the glycine peptide at a 1:1 ratio, and the experimentally determined poses of the respective ligands were extracted from the crystal structures. Parameterization of compounds was performed using the antechamber module of Amber24 with the GAFF2 forcefield⁷⁶, using the AM1-BCC charge model⁷⁷ and a path length of 30. The final bound ABL1 complexes, and the compounds alone, were prepared using tleap from Amber24. The systems were solvated using the TIP3P water model⁷⁸ with the Amber19_SB forcefield⁷⁹, and neutralized with Na⁺ and Cl^– ions in an isometric octahedral box with a minimum distance of 15 Å from the nearest edge. Ion concentrations were subsequently adjusted to mimic experimental buffer conditions (0.14 M K⁺, 0.01 M Na⁺ and 0.15 M Cl⁻).

Simulations were performed stepwise. First, each system was energy-minimized for 2,500 steps using the steepest-descent method, followed by 2,500 steps of conjugate gradient with constraints (100 kcal mol^–1 Å^–2) on all non-solvent atoms. Second, the constraints were released on all atoms, and the system was further minimized for 10,000 steps each of the steepest-descent and conjugate gradient methods. The systems were heated to 300.0 K over 100 ps with a Langevin thermostat in the NVT ensemble with constraints (50 kcal mol^–1 Å^–2) applied on non-solvent atoms. Pressure was maintained at 1 atm with an MC barostat, with a collision frequency of 5 ps⁻¹. Third, positional restraints were slowly released and densities were equilibrated over the course of 500 ps in the NPT ensemble at a temperature of 303.15 K and pressure conditions of 1 atm. Finally, unrestrained production simulations in the NPT ensemble were run for 500 ns for each system in triplicate, 1.5 μs total per compound. All calculations were carried out using an integration step of 2 fs. The SHAKE algorithm was applied to all hydrogen-containing bonds. MD simulations were conducted using the pmemd engine, with CUDA acceleration⁸⁰.

For each production trajectory, the AmberTools cpptraj module⁸¹ was used to calculate root mean square deviation and RMSF values to monitor system equilibration and to assess local flexibility of the compounds, respectively. Atomwise RMSF maps for all compounds, generated as in Fig. 1f, are provided in Supplementary Data 2.

Triplicate simulations were initially performed for each compound. During some simulations, one of the affinity ligands was observed to exit its binding site. This occurred for one of the three initial simulations for each of PonatiLink-1-PEG12, PonatiLink-2-PEG13 and PonatiLink-2-PEG17 (3 out of 33 total initial simulations). These simulations were excluded and additional simulations were performed to replace them, with the goal of comparing linker behaviour while both ligands remained bound. For PonatiLink-2-PEG13 and PonatiLink-2-PEG17, the first subsequent simulation did not demonstrate unbinding; therefore, the data presented are from simulation runs 1, 2 and 4. For PonatiLink-1-PEG12, three additional simulations failed to produce 0.5 μs of sustained binding behaviour; therefore, data presented are from only the first two initial simulations. The total simulation time across systems was 19 μs, with 16 μs of non-excluded simulation time. The entropy of the bound bitopic compounds was calculated using quasi-harmonic analysis⁸² in cpptraj at a temperature of 300 K, using only the compound atoms.

Correlations between mean linker RMSF and entropy per atom across simulations were assessed using Pearson’s correlation coefficient (two-sided test) and reported as the coefficient of determination (R²). Chai-1 (ref. ⁸³) was used to predict the binding pose of RO7304898 shown in Fig. 2a.

Recombinant EGFR expression

EGFR kinase domain (residues 666–1,022) was codon-optimized, synthesized by Twist Bioscience and cloned into the pFastBac vector using the Gibson Assembly method, which produced an amino-terminal 6×His tag and TEV protease cleavage site (ENLYFQG). The entire construct sequence was as follows: MSYYHHHHHHDYDIPTTENLYFQGAMGEAPNQALLRILKETEFKKIKVLGSGAFGTVYKGLWIPEGEKVKIPVAIKELREATSPKANKEILDEAYVMASVDNPHVCRLLGICLTSTVQLITQLMPFGCLLDYVREHKDNIGSQYLLNWCVQIAKGMNYLEDRRLVHRDLAARNVLVKTPQHVKITDFGLAKLLGAEEKEYHAEGGKVPIKWMALESILHRIYTHQSDVWSYGVTVWELMTFGSKPYDGIPASEISSILEKGERLPQPPICTIDVYMIMVKCWMIDADSRPKFRELIIEFSKMARDPQRYLVIQGDERMHLPSPTDSNFYRALMDEEDMDDVVDADEYLIPQQG.

His-TEV-tagged protein was captured using Ni-NTA resin (Thermo Fisher, 88222, 2 ml slurry per litre of culture) at 4 °C for 1 h with constant end-to-end mixing. The loaded beads were then washed with lysis buffer (20 mM Tris pH 8, 500 mM NaCl, 5% glycerol, 1 mM TCEP and 20 mM imidazole) and the protein was eluted with elution buffer (20 mM HEPES pH 8.0, 500 mM NaCl, 5% glycerol and 300 mM imidazole). His-tagged TEV protease (0.025 mg TEV per mg EGFR protein) was then added to the protein solution. The mixture was diluted 1:10 v/v with no salt buffer (20 mM HEPES pH 8.0 and 5% glycerol) and further purified by anion exchange chromatography (HiTrapQ column, Cytiva, 17115301) using a NaCl gradient of 50 mM to 500 mM in no salt buffer. Fractions containing pure EGFR protein were pooled, concentrated and flash-frozen in liquid nitrogen.

Detection of covalent modification of EGFR by whole-protein mass spectrometry

Test compounds were prepared as 100× stock solutions in DMSO. Compounds were diluted with SEC buffer (20 mM HEPES pH 8.0, 150 mM NaCl and 1 mM MgCl₂) to prepare a 2× stock. Recombinant EGFR kinase domain protein was diluted with SEC buffer to 2 μM. Compound solution was mixed with the EGFR protein solution 1:1 (v/v), giving final concentrations of 1 μM EGFR kinase domain and 10 μM compound. The extent of modification was assessed by electrospray mass spectrometry using an Agilent Technologies 6545XT AdvanceBio LC/Q-TOF coupled to an Agilent 1290 Infinity II LC HPLC equipped with a ZORBAX RRHD Eclipse Plus C18 1.8 μm column (Agilent, 959757-902). The mobile phase was a linear gradient of 10–98% acetonitrile–water and 0.05% formic acid. Injection time stamps were used to calculate elapsed time. Time courses were fitted to a one-phase association model in GraphPad Prism (v11.0.2).

Cell viability assays

Cells were seeded on white, opaque, tissue-culture treated 96-well plates (Greiner Bio-One 655083 or Corning 3917) in the central 60 wells of the plate, with border wells filled with 200 μl sterile PBS or water. Cells were seeded in 90 μl complete medium, then allowed to recover overnight. Cells were treated in triplicate with 10 μl of 10× stocks of serial dilutions of compounds (10 conditions including DMSO control, changing pipette tips with each dilution), with a single pipette up and down to mix the suspension cells. In the case of A549 and H1975 cells, compound dilutions were prepared with a Tecan D300e liquid dispenser. Cells were incubated for 3 days before collection. Cell viability was assessed using CellTiter-Glo assays (Promega, G7572). Plates were brought to room temperature by placing on the benchtop, unstacked, for 15 min, then 100 μl CellTiter-Glo reagent diluted 5-fold with PBS was added to each well. Plates were shaken on an orbital shaker at 360 rpm for 20 min, then luminescence was recorded on a Tecan Spark plate reader (luminescence preset, 100 ms of integration time). Luminescence responses were normalized to the mean of the triplicate DMSO control wells for each compound.

Serial dilutions of compounds were prepared on 96-well tissue-culture-treated plates with one row per dilution series. Rows were prepared with one well of 297 μl complete medium and 3 μl DMSO stock (1,000× desired initial concentration), followed by 9 wells of complete medium and 1% DMSO at a volume V depending on the desired serial dilution ratio such that the serial dilution would bring each well to 300 μl. Then (300 μl – V) was serially diluted, pipetting thoroughly to mix, into all wells except the final well, which was left as a DMSO-only control. For example, for a 1,000-fold total dilution ratio, the required serial dilution ratio is \(\sqrt[8]{1,000}\) ≈ 2.371, so (300 – V) = 300/2.371 = 126.5 μl, and V = 173.5 μl. Pipette tips were replaced for each dilution to prevent overtreatment of low-concentration dilutions by compounds adhering to pipette tips.

All K562 cells were seeded at 1,000 cells per well. HUVECs (ATCC, CRL-1730) were seeded at 4,000 cells per well after a pilot experiment (data not shown) indicated that this seeding density produced subsaturated growth in the untreated condition over the experimental time course.

Saturation mutagenesis

BCR::ABL1 cDNA was cloned downstream of EGFP in a pUltra (Addgene, 24129) lentiviral vector by GenScript to produce pUltra wild-type BCR::ABL1 (Addgene, 210432). Twist Bioscience generated a saturating mutagenesis (SM) library of single amino-acid changes in the ABL1 SH3, SH2 and kinase domains (residues 64–512). The SM library was constructed in pools of eight amino acids (8-mers). Stable chemically competent Escherichia coli cells (NEB, C3040I) were transformed with each 8-mer of the SM library, with a coverage of >1,000×, and plated on 15-cm LB agar plates with ampicillin (100 μg ml^–1). Coverage was calculated by estimating colony-forming units across the surface of the plate. Plates were incubated for 48 h at 30 °C, then colonies were scraped off the agar and plasmid DNA was extracted using an EZNA Plasmid DNA Midi kit (Omega Bio-Tek). HEK293T cells at 65–75% confluence in 10-cm plates were transfected with 35 μg of the SM ABL library and 10 μg of third-generation lentiviral packaging plasmids (1:1:1:1, VSV-G, GAG-POL, REV and TAT) using Thermo Fisher Lipofectamine 3000 according to the manufacturer’s protocol (ratio of 5 Lipofectamine: 1 DNA). The next day, the medium was changed to 10 ml fresh RPMI (Cytiva, SH30027.02). After 36 h, the medium was filtered and concentrated using Lenti-X Concentrator (Takara, 631232). Concentrated virus and 6 μg ml^–1 polybrene were added to 10-cm plates containing 5 × 10⁶ K562 cells in 10 ml RPMI (multiplicity of infection < 1). Infection efficiencies ranged from 8 to 15%, with all samples having a library coverage of >1,000×. Infected K562 cells were allowed to recover for 36 h before enrichment by FACS on eGFP at the Penn State Flow Cytometry Core Facility using a Bigfoot Cell Sorter (Thermo Fisher). Sorted cells were expanded for 5 days, then cryopreserved for subsequent experimentation.

Mutational diversity of the K562 BCR::ABL1 mutant library was determined by tile-based amplification. The SH3, SH2 and kinase domains of BCR::ABL1 cDNA were divided into 15 150-bp overlapping tiles. Each tile was PCR-amplified using 1 μg of library gDNA per reaction (30 cycles: 15 s 98 °C, 30 s 65 °C, 30 s 72 °C) with two technical replicates per tile. All PCR reactions were performed using Watchmaker Genomics Equinox master mix. Indexed libraries were pooled at equimolar concentrations and sequenced using paired-end 150 bp reads on the Illumina NovaSeq X Plus platform by Novogene, generating over 1 × 10⁶ paired-end reads per sample. Sequencing data were processed by overlapping, error-correcting and merging paired-end reads using PEAR (v0.9.11)⁸⁴. Merged reads were aligned to the ABL coding sequence (NM_005157.6) using bwa-mem2 (ref. ⁸⁵) in a Linux environment. Variant calling and annotation were performed with a custom R pipeline, generating counts and sequencing depth for each unique mutant at each amino acid position. Subregions and subsequent full-sequence BCR::ABL1 regions for each condition were reconstructed from tiled data using custom Python scripts based on tile position cutoffs. All relevant code is available at GitHub (https://github.com/atlas-biotech/ab_shokat_ponatilink_2).

K562 cells stably expressing the BCR::ABL1 SM library were seeded at a density of 0.15 × 10⁶ cells per ml in 30 ml RPMI (Cytiva), with 10% FBS (Corning) and 1% PSG (Corning) per 15-cm non-treated tissue culture plate (library coverage >500×). Next, 10 mM stock solutions of compounds in DMSO were validated using a CellTiter-Glo 2.0 susceptibility assay before library experiments. Cell pools were treated with compounds for 32 days. Compounds were re-administered to each condition approximately every 3 days for the duration of the experiment by pelleting cells (350g for 5 min) and resuspending in fresh medium with freshly diluted compound. Cell counts and viability were monitored by flow cytometry every 1–3 days for the duration of the experiment using a BD Accuri C6. When live cell density exceeded 0.5 × 10⁶ cells per ml, cell conditions were split to 0.15 × 10⁶ cells per ml.

Immunoblotting

K562 cells were seeded in 6-well tissue-culture-treated plates (Corning 353046) at 1 × 10⁶ cells per well in 2 ml complete medium, and H1975 cells were plated at 0.4 × 10⁶ cells per well in 2 ml complete medium. Cells were allowed to recover at 37 °C overnight. For K562 cell signalling inhibition assays, serial dilutions of 10× compound stocks were prepared as for the cell viability assays, 200 μl of 10× stocks were added to plates, plates were swirled to mix and cells were incubated at 37 °C for 4 h before collection. For H1975 cell signalling inhibition assays, 2 μl of 1,000× DMSO stocks were added directly to plates, plates were swirled to mix and cells were incubated at 37 °C for 24 h before collection. For expression assays, cells were allowed to recover overnight but were not treated. To collect K562 cells, the medium in each well was manually pipetted up and down on each quadrant of the well to detach cells, and the suspension was transferred into a 2 ml tube on ice. Cells were pelleted at 400g for 4 min, then washed twice with 1 ml ice-cold PBS, with pipetting to resuspend. Dry cell pellets were flash-frozen and stored at −80 °C. To collect H1975 cells, medium was aspirated from the plates and cells were washed with 1–2 ml ice-cold PBS in place, then intact plates were stored at −80 °C.

To extract protein from K562 cell pellets, the pellets were thawed on ice for 5 min, then 60 μl lysis buffer (100 mM HEPES pH 7.5, 150 mM NaCl and 0.1% IGEPAL CA-630 (also known as NP-40, Sigma-Aldrich I3021), 1× complete EDTA-free protease inhibitor (Sigma-Aldrich/Roche 11873580001), 1× PhosSTOP phosphatase inhibitor (Sigma-Aldrich/Roche 4906845001)) was added to each pellet and pipetted four times to resuspend. Pellets were incubated on ice for 10 min, then centrifuged at 18,200g for 20 min, and 60 μl of clarified supernatant was transferred to fresh ice-cold PCR strip tubes. To extract protein from H1975 cells, plates were thawed on ice, then 80 μl ice-cold lysis buffer (Pierce RIPA buffer (Thermo 89900) with 1× complete EDTA-free protease inhibitor and 1× PhosSTOP phosphatase inhibitor) was added to each well, incubated for 10–30 min, then scraped with a plastic cell scraper and pipetted into ice-cold 2 ml tubes. Lysates were centrifuged at 18,200g for 15 min, then 72 μl of clarified supernatant was transferred into fresh ice-cold PCR strip tubes.

Protein concentrations were determined using Pierce BCA assays (Thermo Fisher 23225) according to the manufacturer’s instructions, performed in duplicate in 96-well plates with a ratio of 2 μl sample to 98 μl working reagent, and lysates were diluted to 2.5 mg ml^–1 (initial concentrations were generally 3–6 mg ml^–1) with lysis buffer. Gel loading samples were prepared with 5× Laemmli loading buffer (10% w/v SDS, 0.25 M Tris pH 6.8, 0.1% bromophenol blue, 0.5 M dithiothreitol and 50% glycerol) and denatured at 95 °C for 5 min before loading. Next, 10 μl of each loading sample was loaded alongside PageRuler prestained protein ladder (Thermo, 26616) on 4–12% Bis-Tris gels (Invitrogen NuPAGE, for example, WG1403) and separated by electrophoresis at 200 V for 50 min in MOPS buffer (Thermo, NP000102) for blots of K562 cells or for 45 min in MES buffer (Thermo, NP000202) for blots of H1975 cells. Gels of K562 cells were wet-transferred with a Bio-Rad Criterion transfer system to 0.45-μm nitrocellulose membranes (Bio-Rad, 1620115) at 75 V for 60 min at 4 °C with a −20 °C cold pack in Towbin transfer buffer (25 mM Tris, 192 mM glycine, pH 8.6, and 10% methanol) with stirring. Gels of H1975 cells were transferred with an iBlot3 semi-dry transfer system using the Broad Range preset. Subsequent steps were performed with gentle shaking. Membranes were blocked with 5% BSA–TBST (BSA: Sigma-Aldrich, 12659; TBST: 20 mM Tris, 150 mM NaCl, 0.1% Tween-20 and 0.02% sodium azide) for 1 h at room temperature, then cut into strips and probed with 1:1,000 dilutions (except as noted) of primary antibodies in 5% BSA–TBST at 4 °C overnight. Primary antibodies used were as follows, all obtained from Cell Signaling Technology: c-ABL (2862); phospho-c-ABL Y245 (isoform 1b numbering) (2861); STAT5 (94205); phospho-STAT5 Y694 and Y699 (9351); CRKL (3182); phospho-CRKL Y207 (3181); α-tubulin (3873); EGFR (4267); phospho-EGFR Y845 (2231); AKT (4691); phospho-AKT S473 (4060); ERK1/2 (9102); phospho-ERK1/2 T202/Y204 (9101); and BIM (2933). Primary antibody for GAPDH was obtained from Proteintech (60004-1-Ig) and used at 1:5,000 dilution. Antibodies were multiplexed in single solutions as follows: CRKL with phospho-CRKL, ERK with GAPDH and phospho-ERK with GAPDH.

After probing, strips were washed three times briefly with DI water, washed three times (4 min each) with TBST and incubated 1 h at room temperature with a mixture of two secondary antibodies, both at 1:10,000 dilutions in 5% BSA–TBST: LI-COR IRDye 800CW goat anti-rabbit (926-32211) and IRDye 680RD goat anti-mouse (926-68070). Strips were washed three times briefly with DI water, then washed three times (4 min each) with TBST.

For K562 cells, strips from blots were washed briefly 1× with DI water, stored briefly in TBS (TBST without Tween), then imaged simultaneously on a LI-COR Odyssey 9120 imager at 84 μm resolution, medium quality and intensity of 5 on the 800 channel and 3 on the 680 channel. Scans were processed using LI-COR ImageStudio Lite software: full scans were cropped into individual strips, the brightness of cropped strip images was adjusted and images were exported for figures at original resolution. For H1975 cells, strips from blots were imaged individually on a Bio-Rad ChemiDoc MP with the IRDye 800CW and IRDye 680CW settings, using automatic optimal exposure. Scans were processed using Bio-Rad Image Lab software: the brightness of strip images was adjusted and images were exported for figures at original resolution.

In vitro kinase activity assays

In vitro kinase activity assays were performed by Thermo Fisher Scientific (SelectScreen). Experiments were performed in technical duplicate. Data are reported as the per cent inhibition after correction for background fluorescence and normalization to DMSO controls. For dose–response experiments, kinase activity values were calculated (100% – % inhibition) and these values were processed and plotted as for cell viability experiments.

Colony-forming unit assays

Primary peripheral blood mononuclear cells (PBMCs) from healthy individuals were obtained from StemCell Technologies (lot 2410407006) and thawed in DMEM medium (Gibco) supplemented with 20% FBS, 2 mM EDTA, 1% penicillin–streptomycin–l-glutamine (100× PSG, Gibco 10378016) and 500 μg DNase I. Cells were pelleted and resuspended in RPMI supplemented with 10% FBS and 1% PSG and filtered through a 70-μm cell strainer. Around 3 × 10⁴ cells per treatment condition, in triplicate, were incubated with DMSO, ponatinib, asciminib, pona+asc (1:1 ratio) or PonatiLink-2 in 10 nM, 100 nM, 1 μM and 10 μM conditions in human methylcellulose-enriched medium (HSC005, R&D systems). Next, 35-mm culture dishes were inoculated in triplicate with the methylcellulose mixture using 16-gauge non-stick needles. The cultures were incubated at 37 °C and 5% CO₂ for 14 days. Colony counts of BFU-E and CFU-G from the 1 μM and 10 μM conditions were determined by microscopy (Olympus CKX53, ×4 objective) by an individual blinded to the treatment conditions, and counts were normalized for each colony type independently relative to the mean of untreated controls. Following colony counts, plates from all conditions (unstained) were aligned and imaged using a ChemiDoc MP Imaging system (Bio-Rad) under the Coomassie blue setting to visualize BFU-E colony formation. Count data were analysed by negative binomial regression with log link function. Each colony type and dose level was considered separately. Statistics for treatment effects compared with DMSO control were calculated using Dunnett’s method for multiple comparisons (two-sided tests). Results are reported as the log fold change with Dunnett-adjusted P values.

PRISM pooled cell line screen

Ponatinib, asciminib and PonatiLink-2 were tested against the Broad Institute PRISM Laboratory’s MTS026 consortium screen. In brief, cell lines barcoded by lentiviral transduction were subjected to quality control checks, including mycoplasma testing, STR profiling and sequencing each barcode to confirm identity, pooled into groups of 20–25 lines by similar growth rates and cryopreserved. Pools were thawed directly into 384-well plates and treated with compounds at 8 doses in 3-fold dilutions from 2 μM for 5 days. Cells were then lysed with Qiagen TCL buffer and 5 μl of each pool for each cell set was collapsed into one 384-well plate. mRNA was reverse-transcribed, PCR-amplified and quantified by hybridization to Luminex beads as previously described⁵⁸. Data from 865 cell lines passed quality control and were included in analyses. Cell viability as presented in this work was quantified as the Riemann AUC of mean viability readings at each dose, without reference to model fitting. Data-processing details can be found at GitHub (https://github.com/cmap/dockerized_mts).

In vitro kinase bead affinity assays

In vitro KINOMEscan kinase bead affinity assays were performed by Eurofins. Streptavidin-coated magnetic beads were treated with biotinylated small-molecule ligands for 30 min at room temperature to generate affinity resin. The liganded beads were blocked with excess biotin and washed with blocking buffer (SeaBlock (Pierce), 1% BSA, 0.05% Tween 20 and 1 mM DTT) to remove unbound ligand and to reduce nonspecific binding. Binding reactions were assembled by combining kinases, liganded affinity beads and test compounds in 1× binding buffer (20% SeaBlock, 0.17× PBS, 0.05% Tween 20 and 6 mM DTT). Test compounds were prepared as 111× stocks in 100% DMSO. K_d values were determined using an 11-point 3-fold compound dilution series with 3 DMSO control points. All compounds for K_d measurements were distributed by acoustic transfer (non-contact dispensing) in 100% DMSO. The compounds were then diluted directly into the assays such that the final concentration of DMSO was 0.9%. All reactions were performed in polypropylene 384-well plates. Each was a final volume of 0.02 ml. The assay plates were incubated at room temperature with shaking for 1 h and the affinity beads were washed with wash buffer (1× PBS and 0.05% Tween 20). The beads were then re-suspended in elution buffer (1× PBS, 0.05% Tween 20 and 0.5 μM non-biotinylated affinity ligand) and incubated at room temperature with shaking for 30 min. The kinase concentration in the eluates was measured by qPCR. Experiments were performed in technical duplicate.

Dose–response data analysis and visualization

Data from dose–response experiments were analysed and plotted with the ‘doseplotr’ R package, available at GitHub (https://github.com/jackwalkerstevenson/doseplotr). In brief, four-parameter log-logistic functions were fit using the ‘drda’ R package⁸⁶ to the normalized response and the log-molar dose, with the low-dose asymptote constrained to between 80% and 120%. Throughout this work, IC₅₀ refers to the absolute 50% inhibitory concentration, that is, the concentration at which a fitted curve crosses 50% of the control response, and EC₅₀ refers to the relative 50% effective concentration, that is, the concentration at which a fitted curve passes the midpoint between its upper and lower asymptotes. Inhibitory concentrations at other thresholds (for example, IC₉₀) are likewise absolute. This usage distinguishes two definitions of a half-maximal concentration; it is not intended to distinguish inhibitory from non-inhibitory responses.

Animal experiments

The mouse toxicity study was performed by Crown Biosciences in accordance with the Crown Institutional Animal Care and Use Committee (IACUC CBSD-ACUP-001). In brief, 9-week-old female NOD/SCID mice were purchased from the Jackson Laboratory (stock no. 001303) and were housed with ad libitum food and water at 20–26 °C and 30–70% humidity on a 12-h light cycle at the Crown Biosciences San Diego vivarium. Mice were dosed daily with PonatiLink-2-PEG17 100 mg kg^–1 in water 10 ml kg^–1 by intraperitoneal injection. Blood was drawn for compound concentration measurements from the submandibular vein immediately before and 3 days after dosing on day 14, as well as by cardiac bleed at time of euthanasia 7 days after dosing. Blood was processed into plasma and frozen for later analysis. Plasma bioanalysis was performed by liquid chromatography–triple quadrupole mass spectrometry by BioQual Solutions.

Tumour growth studies were performed by the University of California, San Francisco (UCSF) Preclinical Therapeutics Core in accordance with the UCSF Institutional Animal Care and Use Committee (IACUC 194778). In brief, 24 6–7-week-old female NOD/SCID mice per study were purchased from the Jackson Laboratory (stock no. 001303) and were housed with ad libitum food and water on a 12-h light cycle at 20–23 °C and 40–70% humidity at the UCSF Preclinical Therapeutics Core vivarium. Xenografts were established by subcutaneous injection in the right hind flank of 5 × 10⁶ cells in 50 μl each of RPMI 1640 medium and Matrigel. When all tumours reached approximately 100–200 mm³, mice were randomized into 3 groups (n = 8 per group) using a stratified method in Studylog (v4.2.1.3) and dosing was initiated. This day was annotated as day 0. Group sizes were predetermined using power calculations based on an estimated 20% standard deviation and 20–35% effect size, with 8 mice per group selected based on historical data and model complexity. Animal specialists were blinded to treatment group during all procedures and data collection; animal allocation and coding were performed by the Preclinical Therapeutics Core manager. Tumour dimensions and body weight were measured twice a week or once a week as indicated. Tumour volume was calculated to approximate the volume of an ellipsoid as 0.5 × width² × length. The tumour size end point was defined as 20 mm in the maximum dimension, and mice were killed on the same day they were observed to reach end point, according to the IACUC protocol’s Humane End Point standard: “If the greatest tumour dimension exceeds 20 mm in the largest dimension, then the animal will be euthanized.” Mice were also monitored for body weight loss and euthanized if necessary according to the IACUC protocol: “Drug treatment will be stopped if weight loss >15% occurs. If weight loss continues to exceed 15% of pre-tumour implantation body weight, the animal will be euthanized.” Significance of time to end point was calculated using log-rank tests (two-sided, one degree of freedom, test statistic = χ²) between vehicle and each treatment group, corrected for multiple testing by the Holm method.

For the dasatinib + PonatiLink-2 study, K562 cells were used for seeding. Mice were dosed with vehicle (both 5 mM trifluoroacetic acid 5 ml kg^–1 by intraperitoneal injection and 1:1 propylene glycol and water 10 ml kg^–1 by oral gavage), dasatinib (Ambeed, A355193) 5 mg kg^–1 in 1:1 v/v propylene glycol–water 10 ml kg^–1 by oral gavage, or both dasatinib 5 mg kg^–1 in 1:1 v/v propylene glycol–water 10 ml kg^–1 by oral gavage and PonatiLink-2-PEG17 100 mg kg^–1 in water 5 ml kg^–1 by intraperitoneal injection. The vehicle and dasatinib-only groups were dosed daily Monday–Friday for 2 weeks, then Monday–Thursday for one additional week. The dasatinib + PonatiLink-2-PEG17 group received dasatinib on the same schedule, with PonatiLink-2-PEG17 given Thursday–Friday for 2 weeks, then Wednesday–Thursday for one additional week. The study was terminated 3 months after dosing was discontinued. One mouse (477) in the dasatinib-only group was euthanized owing to sudden body weight loss on day 56.

For the ponatinib study, K562 pUltra BCR::ABL1^E255V/T315I cells were used for seeding. Mice were dosed daily Monday–Friday (day 0 was a Thursday) with vehicle (both 5 mM trifluoroacetic acid 5 ml kg^–1 by intraperitoneal injection and 25 mM citrate buffer pH 2.8 10 ml kg^–1 by oral gavage), ponatinib hydrochloride (MCE, HY-108766) 30 mg kg^–1 in 25 mM citrate buffer pH 2.8 10 ml kg^–1 by oral gavage, or PonatiLink-2-PEG17 100 mg kg^–1 in water 5 ml kg^–1 by intraperitoneal injection. Blood was drawn from the saphenous vein for compound concentration measurements from all three groups both immediately before and either 2 h (vehicle and PonatiLink-2) or 6 h (ponatinib) after dosing on day 14, processed into plasma and frozen for subsequent analyses. The study was terminated when all mice in the vehicle-treatment group had reached end point (apart from two spontaneous regressions) on day 42. One mouse (452) in the ponatinib arm was accidentally killed by handling error on day 27. Plasma bioanalysis was performed by liquid chromatography–triple quadrupole mass spectrometry by Oakland Analytics.

The K562 PL2R cell line was cultured from the PonatiLink-2-resistant tumour at the end of the study. The tumour was collected, washed with PBS, minced with scalpels and incubated with 5 ml Accutase (Innovative Cell Technologies) at 37 °C for 45 min with periodic mixing. Accutase was quenched by adding an equal volume of warm complete medium. Dissociated cells and remaining tumour fragments were passed through a 100-μm cell strainer, pressing with a syringe plunger with the strainer immersed in a film of medium, followed by rinsing with additional medium. Strained cells were pelleted, resuspended in complete medium and cultured for 1 week before cryopreserving. Parallel cultures maintained with and without 30 nM PonatiLink-2 for 1 week showed minimal difference in resistance to PonatiLink-2 or parent compounds (data not shown). The cells cultured without supplemental PonatiLink-2 were used for experiments presented here.

Sanger sequencing of spontaneous resistance mutants

gDNA was extracted from 25 mg samples of tumour tissue or pellets of 2 × 10⁶ cultured cells using a DNEasy Blood & Tissue kit (Qiagen 69504). The ABL1 gene was PCR-amplified from the lentiviral integration site using primers spanning from the carboxy-terminus of BCR to the C terminus of the ABL1 kinase domain (forward primer GAAGCTTCTCCCTGACATCCGT; reverse primer TCGTCTTGGTGGGCAGCTC) with thermocycling as follows: 2 min 98 °C, followed by 25 cycles of (10 s 98 °C, 20 s 74 °C, 36 s 72 °C), followed by 2 min 72 °C. The expected 1.7-kb fragment was isolated by gel purification using an EZNA Gel Extraction kit (Omega), eluting with 30 μl nuclease-free water and re-eluting 1× with the eluate for maximum concentration. Sanger sequencing was performed by Elim Biopharm.

Reporting summary

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

The doseplotr R package and shokat_bitopic_scripts repository were written to support this work and used to generate the figures. All code is freely available at GitHub (https://github.com/jackwalkerstevenson/doseplotr (v0.2.0) and https://github.com/jackwalkerstevenson/shokat_bitopic_scripts (v1.0.1)) and Zenodo (https://doi.org/10.5281/zenodo.21751411 and https://doi.org/10.5281/zenodo.21751938, respectively)^88,89. Scripts used to process SM sequencing data are freely available at GitHub (https://github.com/atlas-biotech/ab_shokat_ponatilink_2). Scripts used to process and analyse PRISM data are freely available at GitHub (https://github.com/cmap/dockerized_mts).

1. Gower, C. M., Chang, M. E. K. & Maly, D. J. Bivalent inhibitors of protein kinases. Crit. Rev. Biochem. Mol. Biol. 49, 102–115 (2014).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

2. Braun, T. P., Eide, C. A. & Druker, B. J. Response and resistance to BCR–ABL1-targeted therapies. Cancer Cell 37, 530–542 (2020).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

3. Moslehi, J. J. & Deininger, M. Tyrosine kinase inhibitor-associated cardiovascular toxicity in chronic myeloid leukemia. J. Clin. Oncol. 33, 4210–4218 https://doi.org/10.1200/JCO.2015.62.4718 (2015).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

4. Cohen, P., Cross, D. & Jänne, P. A. Kinase drug discovery 20 years after imatinib: progress and future directions. Nat. Rev. Drug Discov. 20, 551–569 (2021).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

5. Corson, T. W., Aberle, N. & Crews, C. M. Design and applications of bifunctional small molecules: why two heads are better than one. ACS Chem. Biol. 3, 677–692 (2008).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

6. Lane, J. R., Sexton, P. M. & Christopoulos, A. Bridging the gap: bitopic ligands of G-protein-coupled receptors. Trends Pharmacol. Sci. 34, 59–66 (2013).

   Article  CAS  PubMed  Google Scholar 

7. Ricouart, A., Gesquiere, J. C., Tartar, A. & Sergheraert, C. Design of potent protein kinase inhibitors using the bisubstrate approach. J. Med. Chem. 34, 73–78 (1991).

   Article  CAS  PubMed  Google Scholar 

8. Parang, K. et al. Mechanism-based design of a protein kinase inhibitor. Nat. Struct. Biol. 8, 37–41 (2001).

   Article  CAS  PubMed  Google Scholar 

9. Meyer, S. C., Shomin, C. D., Gaj, T. & Ghosh, I. Tethering small molecules to a phage display library: discovery of a selective bivalent inhibitor of protein kinase A. J. Am. Chem. Soc. 129, 13812–13813 (2007).

   Article  ADS  CAS  PubMed  Google Scholar 

10. Békés, M., Langley, D. R. & Crews, C. M. PROTAC targeted protein degraders: the past is prologue. Nat. Rev. Drug Discovery 21, 181–200 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

11. Rodrik-Outmezguine, V. S. et al. Overcoming mTOR resistance mutations with a new-generation mTOR inhibitor. Nature 534, 272–276 (2016).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

12. Burnett, G. L. et al. Discovery of RMC-5552, a selective bi-steric inhibitor of mTORC1, for the treatment of mTORC1-activated tumors. J. Med. Chem. 66, 149–169 (2023).

    Article  CAS  PubMed  Google Scholar 

13. Schram, A. M. et al. The bi-steric, mTORC1-selective inhibitor, RMC-5552, in advanced solid tumors: a Phase 1 trial. Clin. Cancer Res. 31, 4933–4943 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

14. Lou, K. et al. IFITM proteins assist cellular uptake of diverse linked chemotypes. Science 378, 1097–1104 (2022).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

15. Wieduwilt, M. J. Ph⁺ ALL in 2022: is there an optimal approach? Hematology 2022, 206–212 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

16. Greuber, E. K., Smith-Pearson, P., Wang, J. & Pendergast, A. M. Role of ABL family kinases in cancer: from leukaemia to solid tumours. Nat. Rev. Cancer 13, 559–571 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

17. Roberts, K. G. et al. Targetable kinase-activating lesions in Ph-like acute lymphoblastic leukemia. N. Engl. J. Med. 371, 1005–1015 (2014).

    Article  PubMed  PubMed Central  Google Scholar 

18. Druker, B. J. et al. Efficacy and safety of a specific inhibitor of the BCR–ABL tyrosine kinase in chronic myeloid leukemia. N. Engl. J. Med. 344, 1031–1037 (2001).

    Article  CAS  PubMed  Google Scholar 

19. Saussele, S. et al. Discontinuation of tyrosine kinase inhibitor therapy in chronic myeloid leukaemia (EURO-SKI): a prespecified interim analysis of a prospective, multicentre, non-randomised, trial. Lancet Oncol. 19, 747–757 (2018).

    Article  CAS  PubMed  Google Scholar 

20. Hochhaus, A. & Ernst, T. TKI discontinuation in CML: how do we make more patients eligible? How do we increase the chances of a successful treatment-free remission? Hematology Am. Soc. Hematol. Educ. Program 2021, 106–112 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

21. Gorre, M. E. et al. Clinical resistance to STI-571 cancer therapy caused by BCR–ABL gene mutation or amplification. Science 293, 876–880 (2001).

    Article  CAS  PubMed  Google Scholar 

22. O’Hare, T. et al. AP24534, a pan-BCR-ABL inhibitor for chronic myeloid leukemia, potently inhibits the T315I mutant and overcomes mutation-based resistance. Cancer Cell 16, 401–412 (2009).

    Article  PubMed  PubMed Central  Google Scholar 

23. Cortes, J. E. et al. A Phase 2 trial of ponatinib in Philadelphia chromosome-positive leukemias. N. Engl. J. Med. 369, 1783–1796 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

24. Zabriskie, M. S. et al. BCR–ABL1 compound mutations combining key kinase domain positions confer clinical resistance to ponatinib in Ph chromosome-positive leukemia. Cancer Cell 26, 428–442 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

25. Schmitt, M. W. et al. Single-molecule sequencing reveals patterns of preexisting drug resistance that suggest treatment strategies in Philadelphia-positive leukemias. Clin. Cancer Res. 24, 5321–5334 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

26. Soverini, S. et al. BCR–ABL1 compound mutants: prevalence, spectrum and correlation with tyrosine kinase inhibitor resistance in a consecutive series of Philadelphia chromosome-positive leukemia patients analyzed by NGS. Leukemia 35, 2102–2107 (2020).

    Article  PubMed  Google Scholar 

27. Hantschel, O. et al. A myristoyl/phosphotyrosine switch regulates c-Abl. Cell 112, 845–857 (2003).

    Article  CAS  PubMed  Google Scholar 

28. Nagar, B. et al. Structural basis for the autoinhibition of c-Abl tyrosine kinase. Cell 112, 859–871 (2003).

    Article  CAS  PubMed  Google Scholar 

29. Hochhaus, A. et al. Asciminib in newly diagnosed chronic myeloid leukemia. N. Engl. J. Med. 391, 885–898 (2024).

    Article  CAS  PubMed  Google Scholar 

30. Yeung, D. T., Shanmuganathan, N. & Hughes, T. P. Asciminib: a new therapeutic option in chronic-phase CML with treatment failure. Blood 139, 3474–3479 (2022).

    Article  CAS  PubMed  Google Scholar 

31. Leyte-Vidal, A. et al. BCR::ABL1 kinase N-lobe mutants confer moderate to high degrees of resistance to asciminib. Blood 144, 639–645 (2024).

    Article  CAS  PubMed  Google Scholar 

32. Eadie, L. N. et al. Asciminib is a novel inhibitor of ABL1 and ABL2 gene fusions in ALL but requires the ABL SH3 domain for efficacy. Blood 144, 1022–1026 (2024).

    Article  CAS  PubMed  Google Scholar 

33. Leske, I. B. & Hantschel, O. The e13a3 (b2a3) and e14a3 (b3a3) BCR::ABL1 isoforms are resistant to asciminib. Leukemia 38, 2041–2045 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

34. Leyte-Vidal, A. et al. Absence of ABL1 exon 2-encoded SH3 residues in BCR::ABL1 destabilizes the autoinhibited kinase conformation and confers resistance to asciminib. Leukemia 38, 2046–2050 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

35. Wylie, A. A. et al. The allosteric inhibitor ABL001 enables dual targeting of BCR–ABL1. Nature 543, 733–737 (2017).

    Article  ADS  CAS  PubMed  Google Scholar 

36. Eide, C. A. et al. Combining the allosteric inhibitor asciminib with ponatinib suppresses emergence of and restores efficacy against highly resistant BCR–ABL1 mutants. Cancer Cell 36, 431–443 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

37. Lipton, J. H. et al. Ponatinib versus imatinib for newly diagnosed chronic myeloid leukaemia: an international, randomised, open-label, phase 3 trial. Lancet Oncol. 17, 612–621 (2016).

    Article  CAS  PubMed  Google Scholar 

38. Senapati, J. et al. Management of chronic myeloid leukemia in 2023—common ground and common sense. Blood Cancer J. 13, 58 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

39. Eide, C. A. et al. Overcoming clinical BCR–ABL1 compound mutant resistance with combined ponatinib and asciminib therapy. Cancer Cell 42, 1486–1488 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

40. Kim, C. et al. A biophysical framework for double-drugging kinases. Proc. Natl Acad. Sci. USA 120, e2304611120 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

41. Zhou, T. et al. Structural mechanism of the pan-BCR–ABL inhibitor ponatinib (AP24534): lessons for overcoming kinase inhibitor resistance. Chem. Biol. Drug Des. 77, 1–11 (2011).

    Article  CAS  PubMed  Google Scholar 

42. Huang, W.-S. et al. Discovery of 3-[2-(imidazo[1,2-b]pyridazin-3-yl)ethynyl]-4-methyl-N-{4-[(4-methylpiperazin-1-yl)methyl]-3-(trifluoromethyl)phenyl}benzamide (AP24534), a potent, orally active pan-inhibitor of breakpoint cluster region–abelson (BCR–ABL) kinase including the T315I gatekeeper mutant. J. Med. Chem. 53, 4701–4719 (2010).

    Article  CAS  PubMed  Google Scholar 

43. Karaman, M. W. et al. A quantitative analysis of kinase inhibitor selectivity. Nat. Biotechnol. 26, 127–132 (2008).

    Article  ADS  CAS  PubMed  Google Scholar 

44. Vasta, J. D. et al. Quantitative, wide-spectrum kinase profiling in live cells for assessing the effect of cellular ATP on target engagement. Cell Chem. Biol. 25, 206–214 (2018).

    Article  CAS  PubMed  Google Scholar 

45. Kwarcinski, F. E. et al. Conformation-selective analogues of dasatinib reveal insight into kinase inhibitor binding and selectivity. ACS Chem. Biol. 11, 1296–1304 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

46. Johnson, T. K. et al. Synergy and antagonism between allosteric and active-site inhibitors of Abl tyrosine kinase. Angew. Chem. Int. Ed. 60, 20196–20199 (2021).

    Article  ADS  CAS  Google Scholar 

47. Yarden, Y. & Sliwkowski, M. X. Untangling the ErbB signalling network. Nat. Rev. Mol. Cell Biol. 2, 127–137 (2001).

    Article  CAS  PubMed  Google Scholar 

48. Cross, D. A. E. et al. AZD9291, an irreversible EGFR TKI, overcomes T790M-mediated resistance to EGFR inhibitors in lung cancer. Cancer Discov. 4, 1046–1061 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

49. Jia, Y. et al. Overcoming EGFR(T790M) and EGFR(C797S) resistance with mutant-selective allosteric inhibitors. Nature 534, 129–132 (2016).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

50. Heinig, K. et al. Chemical, analytical and pharmacokinetic characterisation of RO7304898, an API consisting of two rapidly interconverting diastereoisomers. Pharm. Res. 39, 653–667 (2022).

    Article  CAS  PubMed  Google Scholar 

51. Beyett, T. S. et al. Molecular basis for cooperative binding and synergy of ATP-site and allosteric EGFR inhibitors. Nat. Commun. 13, 2530 (2022).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

52. Li, Q. et al. Discovery of potent and noncovalent reversible EGFR kinase inhibitors of EGFR^{L858R/T790M/C797S}. ACS Med. Chem. Lett. 10, 869–873 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

53. Wittlinger, F. et al. Design of a “two-in-one” mutant-selective epidermal growth factor receptor inhibitor that spans the orthosteric and allosteric sites. J. Med. Chem. 65, 1370–1383 (2022).

    Article  CAS  PubMed  Google Scholar 

54. Mauro, M. J. et al. Asciminib monotherapy in patients with CML-CP without BCR::ABL1 T315I mutations treated with at least two prior TKIs: 4-year phase 1 safety and efficacy results. Leukemia 37, 1048–1059 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

55. Innes, A. J. et al. Impact of BCR::ABL1 single nucleotide variants on asciminib efficacy. Leukemia 38, 2443–2455 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

56. Shah, N. P. et al. Transient potent BCR–ABL inhibition is sufficient to commit chronic myeloid leukemia cells irreversibly to apoptosis. Cancer Cell 14, 485–493 (2008).

    Article  CAS  PubMed  Google Scholar 

57. Manley, P. W., Barys, L. & Cowan-Jacob, S. W. The specificity of asciminib, a potential treatment for chronic myeloid leukemia, as a myristate-pocket binding ABL inhibitor and analysis of its interactions with mutant forms of BCR–ABL1 kinase. Leuk. Res. 98, 106458 (2020).

    Article  CAS  PubMed  Google Scholar 

58. Corsello, S. M. et al. Discovering the anticancer potential of non-oncology drugs by systematic viability profiling. Nat. Cancer 1, 235–248 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

59. De Braekeleer, E. et al. ABL1 fusion genes in hematological malignancies: a review. Eur. J. Haematol. 86, 361–371 (2011).

    Article  PubMed  Google Scholar 

60. Latifi, Y. et al. Thrombotic microangiopathy as a cause of cardiovascular toxicity from the BCR–ABL1 tyrosine kinase inhibitor ponatinib. Blood 133, 1597–1606 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

61. Cortes, J. E. et al. Final 5-year study results of DASISION: the dasatinib versus imatinib study in treatment-naïve chronic myeloid leukemia patients trial. J. Clin. Oncol. 34, 2333–2340 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

62. Schieven, G. L. et al. Dasatinib once-daily dose regimen provides robust anti-leukemic activity while avoiding suppression of T cell activation. Blood 114, 2198–2198 (2009).

    Article  Google Scholar 

63. Adrián, F. J. et al. Allosteric inhibitors of Bcr–abl–dependent cell proliferation. Nat. Chem. Biol. 2, 95–102 (2006).

    Article  PubMed  Google Scholar 

64. Zhang, J. et al. Targeting Bcr–Abl by combining allosteric with ATP-binding-site inhibitors. Nature 463, 501–506 (2010).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

65. Hornberger, K. R. & Araujo, E. M. V. Physicochemical property determinants of oral absorption for PROTAC protein degraders. J. Med. Chem. 66, 8281–8287 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

66. Krishnamurthy, V. M., Semetey, V., Bracher, P. J., Shen, N. & Whitesides, G. M. Dependence of effective molarity on linker length for an intramolecular protein–ligand system. J. Am. Chem. Soc. 129, 1312–1320 (2007).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

67. Wu, C. & Lu, Y. Inclusion of high molecular weight dextran in calcium phosphate-mediated transfection significantly improves gene transfer efficiency. Cell. Mol. Biol. 53, 67–74 (2007).

    CAS  PubMed  PubMed Central  Google Scholar 

68. Sena-Esteves, M. & Gao, G. Production of high-titer retrovirus and lentivirus vectors. Cold Spring Harb. Protoc. https://doi.org/10.1101/pdb.prot095687 (2018).

69. Chicaybam, L. et al. An efficient electroporation protocol for the genetic modification of mammalian cells. Front. Bioeng. Biotechnol. 4, 99 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

70. Ye, J. et al. Primer-BLAST: a tool to design target-specific primers for polymerase chain reaction. BMC Bioinformatics 13, 134 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

71. Kluesner, M. G. et al. EditR: a method to quantify base editing from Sanger sequencing. CRISPR J. 1, 239–250 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

72. Wu, S. Q. et al. Extensive amplification of bcr/abl fusion genes clustered on three marker chromosomes in human leukemic cell line K-562. Leukemia 9, 858–862 (1995).

    CAS  PubMed  Google Scholar 

73. Virgili, A. & Nacheva, E. P. Genomic amplification of BCR/ABL1 and a region downstream of ABL1 in chronic myeloid leukaemia: a FISH mapping study of CML patients and cell lines. Mol. Cytogenet. 3, 15 (2010).

    Article  PubMed  PubMed Central  Google Scholar 

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

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

75. Sali, A. & Blundell, T. L. Comparative protein modelling by satisfaction of spatial restraints. J. Mol. Biol. 234, 779–815 (1993).

    Article  CAS  PubMed  Google Scholar 

76. Wang, J., Wolf, R. M., Caldwell, J. W., Kollman, P. A. & Case, D. A. Development and testing of a general amber force field. J. Comput. Chem. 25, 1157–1174 (2004).

    Article  ADS  CAS  PubMed  Google Scholar 

77. Jakalian, A., Jack, D. B. & Bayly, C. I. Fast, efficient generation of high-quality atomic charges. AM1-BCC model: II. Parameterization and validation. J. Comput. Chem. 23, 1623–1641 (2002).

    Article  ADS  CAS  PubMed  Google Scholar 

78. Jorgensen, W. L., Chandrasekhar, J., Madura, J. D., Impey, R. W. & Klein, M. L. Comparison of simple potential functions for simulating liquid water. J. Chem. Phys. 79, 926–935 (1983).

    Article  ADS  CAS  Google Scholar 

79. Case, D. A. et al. The Amber biomolecular simulation programs. J. Comput. Chem. 26, 1668–1688 (2005).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

80. Salomon-Ferrer, R., Götz, A. W., Poole, D., Le Grand, S. & Walker, R. C. Routine microsecond molecular dynamics simulations with AMBER on GPUs. 2. Explicit solvent particle mesh Ewald. J. Chem. Theory Comput. 9, 3878–3888 (2013).

    Article  CAS  PubMed  Google Scholar 

81. Roe, D. R. & Cheatham, T. E. PTRAJ and CPPTRAJ: software for processing and analysis of molecular dynamics trajectory data. J. Chem. Theory Comput. 9, 3084–3095 (2013).

    Article  CAS  PubMed  Google Scholar 

82. Chang, C.-E., Chen, W. & Gilson, M. K. Evaluating the accuracy of the quasiharmonic approximation. J. Chem. Theory Comput. 1, 1017–1028 (2005).

    Article  CAS  PubMed  Google Scholar 

83. Chai Discovery. Chai-1: decoding the molecular interactions of life. Preprint at bioRxiv https://doi.org/10.1101/2024.10.10.615955 (2024).

84. Zhang, J., Kobert, K., Flouri, T. & Stamatakis, A. PEAR: a fast and accurate Illumina paired-end read merger. Bioinformatics 30, 614–620 (2014).

    Article  CAS  PubMed  Google Scholar 

85. Li, H. Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM. Preprint at https://doi.org/10.48550/arXiv.1303.3997 (2013).

86. Malyutina, A., Tang, J. & Pessia, A. drda: an R package for dose–response data analysis using logistic functions. J. Stat. Softw. 106, 1–26 (2023).

    Article  Google Scholar 

87. Stevenson, J. et al. Source data for “A design approach for bitopic kinase inhibitors” [Dataset]. Zenodo https://doi.org/10.5281/zenodo.21752810 (2026).

88. Stevenson, J. jackwalkerstevenson/doseplotr: doseplotr v0.2.0 (Version v0.2.0) [Computer software]. Zenodo https://doi.org/10.5281/zenodo.21751411 (2026).

89. Stevenson, J. & Peterson, S. Y. J. jackwalkerstevenson/shokat_bitopic_scripts: shokat_bitopic_scripts v1.0.1 for publication (Version v1.0.1) [Computer software]. Zenodo https://doi.org/10.5281/zenodo.21751938 (2026).

Download references

Author notes

1. These authors contributed equally: Jack W. Stevenson, Kevin Lou

Authors and Affiliations

1. Department of Cellular and Molecular Pharmacology, University of California, San Francisco, San Francisco, CA, USA

   Jack W. Stevenson, Kevin Lou, Sky Y. J. Peterson, Anastasiia Grynyova, Sydney M. Figueroa & Kevan M. Shokat

2. Howard Hughes Medical Institute, University of California, San Francisco, San Francisco, CA, USA

   Jack W. Stevenson, Kevin Lou, Anastasiia Grynyova, Sydney M. Figueroa & Kevan M. Shokat

3. Department of Bioengineering and Therapeutic Sciences, University of California, San Francisco, San Francisco, CA, USA

   Kevin Lou, Kenneth Huang, Anastasiia Grynyova, Sydney M. Figueroa, Ignacia Echeverria & Andrej Šali

4. Quantitative Biosciences Institute, University of California, San Francisco, San Francisco, CA, USA

   Kevin Lou, Ignacia Echeverria & Andrej Šali

5. Department of Biomedical Engineering, The Pennsylvania State University, University Park, PA, USA

   Joshua A. Reynolds, Haider Inam & Justin R. Pritchard

6. Department of Pharmacological Sciences, Stony Brook University, Stony Brook, NY, USA

   Ian R. Outhwaite, Emma L. Gebauer & Markus A. Seeliger

7. Medical Scientist Training Program, Stony Brook University, Stony Brook, NY, USA

   Ian R. Outhwaite

8. Huck Institutes of the Life Sciences, The Pennsylvania State University, University Park, PA, USA

   Ivan Sokirniy & Justin R. Pritchard

9. Section of Cardio-Oncology and Immunology, Department of Medicine, University of California, San Francisco, San Francisco, CA, USA

   Amir Z. Munir & Javid J. Moslehi

10. Department of Chemistry, University of California, Berkeley, Berkeley, CA, USA

    Celine D. Wang & Ziyang Zhang

11. Department of Medicine, Division of Hematology and Oncology, University of California, San Francisco, San Francisco, CA, USA

    Ariel Leyte-Vidal & Neil P. Shah

12. Department of Biochemistry and Molecular Biology, University of Miami Miller School of Medicine, Miami, FL, USA

    Ariel Leyte-Vidal

13. Broad Institute of MIT and Harvard, Cambridge, MA, USA

    Matthew G. Rees, Melissa M. Ronan & Jennifer A. Roth

14. Helen Diller Family Comprehensive Cancer Center, University of California, San Francisco, San Francisco, CA, USA

    Fernando Salangsang, Yan Zeng & Veronica Steri

15. Department of Pharmaceutical Chemistry, University of California, San Francisco, San Francisco, CA, USA

    Andrej Šali

16. Department of Molecular and Cellular Biology, University of California, Berkeley, Berkeley, CA, USA

    Ziyang Zhang

Contributions

K.L. and K.M.S. conceived the project. J.W.S. led the project. K.L. synthesized DasatiLink compounds, OsimertiLink compounds, PonatiLink-1-PEG20 and NanoBRET fluorescent tracers. J.W.S. synthesized PonatiLink-1 and PonatiLink-2 compounds. S.Y.J.P. synthesized PonatiLink-1 compounds. I.S. prepared the K562 cell lines. J.W.S., I.S., K.L. and S.Y.J.P. performed the K562 cell viability experiments. I.R.O. and E.L.G. performed the NanoBRET experiments. K.H. performed the MD simulations. C.D.W. prepared EGFR protein and performed the labelling experiments. A.G. and S.M.F. performed the H1975 and A549 cell viability experiments and immunoblot experiments. J. A. Reynolds and H.I. performed the saturation mutagenesis experiments, and J. A. Reynolds analysed the data. J.W.S. performed the K562 immunoblot experiments. A.L.-V. performed the PBMC colony-formation assays. M.G.R., M.M.R. and J. A. Roth performed the PRISM cell line screen. A.Z.M. and J.W.S. performed the HUVEC cell viability experiments. J.W.S. and V.S. designed the mouse experiments. F.S. and Y.Z. performed the tumour growth experiments. J.W.S. cultured and analysed the resistant tumour. J.W.S. wrote the doseplotr analysis software and drafted the manuscript and the figures. K.M.S. and K.L. supervised and interpreted the synthesis, cell viability and K562 immunoblot experiments. N.P.S. supervised and interpreted the colony-formation assays. I.E. and A.Š. supervised and interpreted the MD work. Z.Z. supervised and interpreted the EGFR kinetics work. J.J.M. supervised and interpreted the HUVEC cell viability work. M.A.S. supervised and interpreted the NanoBRET work. J.R.P. supervised and interpreted the K562 cell line preparation, cell viability and saturation mutagenesis work. All authors reviewed the manuscript.

Corresponding authors

Correspondence to Kevin Lou or Kevan M. Shokat.

Competing interests

J.W.S., K.L. and K.M.S. are inventors on patent WO2023114759A2 related to DasatiLink-1, PonatiLink-1 and PonatiLink-2 and owned by the UC Regents. Parents of K.L. hold equity in and are employed by Pharmaron. K.M.S. owns stock and/or receives monetary compensation from Black Arrow, BridGene Biosciences, Erasca, Exai, G Protein Therapeutics, Genentech-Roche, Kumquat Biosciences, Kura Oncology, Lyterian, Merck, Montara Therapeutics, Nextech, Revolution Medicines, Rezo, Tahoe, Totus, Type6 Therapeutics, Vx Capital and Wellspring Biosciences (Araxes Pharma). J. A. Reynolds and H.I. are founders of and hold equity in Atlas Biotech. J.R.P. has consulted for Theseus Pharmaceuticals, MOMA Therapeutics, Takeda Pharmaceuticals, Curie.Bio, Galapagos Pharmaceuticals and WuXi Nextcode. J.R.P. holds equity in Theseus Pharmaceuticals and MOMA Therapeutics and has received research funding from Theseus Pharmaceuticals. J.J.M. has provided consulting or advisory roles for Pfizer, Novartis, Bristol-Myers Squibb, Deciphera, Takeda, AstraZeneca, Regeneron, Myovant, Silverback Therapeutics, Kurome Therapeutics, Kiniksa Pharmaceuticals, Daiichi Sankyo, CRC Oncology, BeiGene, Prelude Therapeutics, TransThera Sciences, Antev, IQVIA, AskBio, Labcorp, Paladin, Quell Therapeutics, Voyager Therapeutics, Bitterroot Bio, Repare Therapeutics, Teva and Cytokinetics. N.P.S. has received clinical research funding from Kumquat Biosciences.

Peer review information

Nature thanks Nathanael Gray, Charles Sawyers and the other, anonymous, reviewer for their contribution to the peer review of this work. Peer reviewer reports are available.

Source Data Fig. 1, 2, 3, 4, 5 and Source Data Extended Data Fig. 2, 3, 4, 5, 7, 8, 9, 10 (download ZIP )

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