Discovery of a bifunctional PKMYT1-targeting PROTAC empowered by AI-generation Authors: Yazhou Wang, Xiaomin Wang, Tingting Liu, Chao Wang, Qingshuo Meng, Fanye Meng, Jiaojiao Yu, Jinxin Liu, Yaya Fan, David Gennert, Frank W Pun, Alex Aliper, Feng Ren, Man Zhang, Xin Cai, Xiao Ding, Alex Zhavoronkov Published: 2025 ABSTRACT AbstractPKMYT1 has recently emerged as a compelling therapeutic target for precision cancer therapy due to its synthetic lethality with oncogenic alterations such as CCNE1 amplification and mutations in FBXW7 and PPP2R1A. Current small molecule PKMYT1 inhibitors face limitations, such as insufficient molecular diversity and poor selectivity. We herein use our generative AI platform to develop a bifunctional PKMYT1 degrader by linking an entirely novel PKMYT1 inhibitor to an optimized cereblon (CRBN) binder. The lead PROTAC D16-M1P2 demonstrates dual mechanisms of PKMYT1 degradation and inhibition, with strong antiproliferative potency facilitated by high selectivity. It also exhibits favorable oral bioavailability, stronger pharmacodynamic effects relative to the PKMYT1 inhibitor alone, and robust antitumor response as a monotherapy in xenograft models. This PROTAC serves as a precise chemical probe to explore PKMYT1 biology and a promising lead for further cancer therapy exploration.Subject terms: Computational biophysics, Computational models, X-ray crystallography, Drug discovery and development Introduction Cell cycle regulation is frequently disrupted in cancer, leading to uncontrolled cell proliferation, genomic instability, and ultimately tumor progression1, but also to cancer-specific vulnerabilities2. Targeting synthetically lethal genetic interactions offers potential therapeutic avenues to indirectly target proteins previously considered “undruggable”3, such as highly structurally similar cell cycle regulatory proteins4,5, while sparing normal cells and minimizing side effects6. PKMYT1, a cytoplasmic membrane-associated serine-threonine protein kinase that regulates CDK1 and the G2/M checkpoint7, has recently emerged as a synthetically lethal target for cancers with CCNE1 (cyclin E1) amplification and deleterious mutations in FBXW7 and PPP2R1A8,9. CCNE1 amplification or overexpression increases replication stress and necessitates PKMYT1-mediated inactivating phosphorylation of CDK1, thereby preventing premature mitotic entry prior to the completion of genome duplication and subsequent catastrophic genomic instability10. Selective inhibitors of CDK2, the downstream effector directly activated by CCNE1, have faced challenges due to suboptimal selectivity and potential resistance10,11. PKMYT1 inhibition orthogonally disrupts the G2/M checkpoint and induces synthetic lethality in the context of CCNE1 amplification or overexpression12, which may circumvent both the toxicity and resistance risks associated with CDK2-targeted therapies. Additionally, loss-of-function mutations in FBXW7, an E3 ubiquitin ligase that mediates cyclin E degradation, and in PPP2R1A, a PP2A phosphatase subunit involved in cell cycle regulation, induce dependence on CDK1 inhibition and are promising genetic backgrounds for PKMYT1 inhibition13–15. The first-in-class PKMYT1 inhibitor RP-63068 and other inhibitors based on its structure feature an inner- or outer-cyclic amide as the PKMYT1 hinge-domain binder16. These inhibitors face dose-limiting toxicity and limited clinical benefit9, as well as concerns of acquired resistance mutations17,18. Furthermore, PKMYT1 has been reported to physically interact with β-catenin, non-catalytically blocking recognition by the E3-ligase SCFβ-TrCP, thereby stabilizing β-catenin and constitutively activating Wnt signaling, which in turn enhances tumorigenesis19, suggesting that combined targeting of both catalytic and non-catalytic functions of PKMYT1 may improve antitumor efficacy. Proteolysis targeting chimeras (PROTACs)20–22 represent a method of protein inactivation that may avoid these pitfalls. PROTACs consist of two distinct chemical moieties: one binds the protein of interest (POI), while the other binds a cellular E3-ubiquitin ligase subunit, such as cereblon (CRBN)23. PROTACs induce proximity between the POI and the E3 ligase, thereby facilitating target protein degradation via the ubiquitin/proteasome system24. The unique and plastic POI-E3 ligase interaction interfaces and the tissue-selective nature of E3 ligases can improve selectivity over the POI ligand itself25–28. The event-driven pharmacology of PROTAC-mediated degradation eliminates the need for long residence time and saturated drug concentration. In this work, we employ the multi-modal AI-driven Chemistry42 platform to generate a novel POI-binder chemotype with potent inhibition against PKMYT1. Following comprehensive CRBN binder screening and linker optimization, we incorporate the novel PKMYT1 inhibitor into the D16-M1P2 PROTAC, a PKMYT1 degrader. This degrader demonstrates robust and highly specific degradation activity synergistic with CCNE1/FBXW7/PPP2R1A alterations. D16-M1P2 exhibits a dual mechanism of action, encompassing both PKMYT1 degradation and inhibition, exhibiting stronger and sustained pharmacodynamic (PD) effects relative to the foundational POI inhibitor used to initiate design. Our findings provide a tool for studying PKMYT1 biology and offering a promising class of drugs to counteract PKMYT1 activity in cancer. Results Discovery of a novel PKMYT1 inhibitor by AI-driven generation via Chemistry42 The PKMYT1 inhibitor RP-6306 bears five H-bond donors (HBDs) located in the deep pocket of the PKMYT1 ATP-binding site, making it suboptimal as a POI ligand29. The soaked structure of PKMYT1 in complex with RP-6306 (PDB code: 8D6E) depicts RP-6306 with its primary amino group engaging in hydrogen bonding with the gatekeeper residue Thr187 and hinge residue Glu188, its carboxamide group forming hydrogen bonds with Cys190 and Gly191 in the hinge region, its hydroxyl of the phenol group forming a water-mediated hydrogen bond with Glu157, Phe252, and His161, and its pyridine nitrogen forming a water-mediated hydrogen bond with the side chain of Lys139 (Supplementary Fig. 1a). To further examine the dynamics and stability of these interactions, a 200-ns molecular dynamics (MD) simulation was conducted, identifying highly persistent and stable hydrogen bonding interactions (Supplementary Fig. 1b and 1c). Specifically, hydrogen bonding between the primary amino group and Thr187 remains stable, with an occupancy of 88.4%, while interaction with Glu188 is sustained at 22.2%. The carboxamide group shows strong hydrogen bonding with Cys190 (80.5%) and Gly191 (97.4%), reaffirming its role as a key structural feature contributing to binding stability. MD simulations further reveal additional hydrophobic and van der Waals interactions that complement the hydrogen bonding network, such as interactions with Val124 and Phe240. These secondary interactions likely contribute to stabilizing RP-6306 within the PKMYT1 binding pocket. The consistency of these interactions across the 200 ns simulation trajectory underscores the robustness of the binding mode of RP-6306 and its critical engagement with the hinge and key active site residues of PKMYT1, providing a valuable framework for designing and identifying novel inhibitors that replicate these critical binding features. Unlike RP-6306, the approved broad-spectrum kinase inhibitor dasatinib (PKMYT1 IC50 63 nM) extends into the solvent-exposed region and forms a hydrogen bond with residue Gln196 (Supplementary Fig. 1d). Despite the absence of a phenol hydroxyl group in dasatinib, a structural water molecule in the back pocket maintains stable hydrogen bonds with surrounding residues, suggesting that PKMYT1 inhibitors can extend from two conserved water molecule interactions in the back pocket of the hinge region into the solvent region (Supplementary Fig. 1e). We thus utilized the combined features of RP-6306 and dasatinib for molecule design. The complex structure of PKMYT1 with RP-6306 (PDB: 8D6E) served as a starting point for structural generation. The 2,4-dimethyl-3-phenol group of RP-6306 in the back pocket was anchored and kept constant. The pharmacophores of RP-6306 in the hinge region and dasatinib in the solvent area were combined into a new pharmacophore model, which was then used as the input model for AI-driven generation using the Chemistry42 platform30 (Fig. 1a, upper left). The strategy involved fragment replacement in the hinge region and directed fragment growing in the solvent area, using over 40 AI-driven generative models. Internal parameters such as protein-ligand interactions, 3D shape compatibility with the binding pocket, pocket volume occupation, and molecular novelty were applied as filters (Supplementary Fig. 1f). A total of 2023 molecules were generated and subsequently subjected to clustering, re-docking, and visual inspection to evaluate their protein-ligand interactions. In addition, these re-docked molecules were broken down into fragments, with particular emphasis on analyzing the hinge region scaffolds and solvent-exposed groups. Building upon the prior analysis of protein-ligand interactions, novel fragments with favorable binding characteristics were identified. Functional groups corresponding to those potentially favorable fragments (Fig. 1a, upper middle) were systematically incorporated into novel molecules, followed by docking simulations and visual evaluations. AI-driven generation via Chemistry42 platform and development process of novel chemotype of PKMYT1 inhibitors and PROTAC. a AI-generation and optimization resulting in compound 4. Biochemical binding affinity (HTRF assay) and cellular pCDK1 at Thr14 (in-cell western assay) denote IC50 values. The largest red dot denotes the PKMYT1 in the KinomeScan map tested at 1000 nM. All the data are presented as mean of at least two duplicates. CLp clearance in plasma. b X-ray co-crystal structure of compound 2 with PKMYT1. c Predicted binding pose of compound 3. d Predicted binding pose of compound 4. e The modeled PKMYT1-CRL4CRBN complex. The green cartoon represents PKMYT1, with the predicted ubiquitinated Lys345 depicted by green sticks. f Structure of compound D1. g Comparison of pCDK1 inhibitory activity of PROTAC D1 with the PKMYT1-binding inhibitor component compound 4 in HCC1569 cells. Data are presented as the mean (n = 2 for bioactivity and n = 3 for PK). Source data are provided as a Source Data file. Ultimately, one molecule (compound 1) was selected for synthesis based on predicted interactions and overall potential, taking into account synthetic accessibility and drug-like properties. Subsequently, the Golden Cubes toolkit within the Chemistry42 platform, developed to streamline the early discovery process of kinase inhibitors and predict their potential off-target effects, was utilized to evaluate the off-target probabilities of these novel molecules. As a result, compared to dasatinib and RP-6306, compound 1 exhibited a significantly lower possibility of off-target interactions with ABL1, SRC, LCK, and JAKs (Supplementary Fig. 2a). Compound 1 (Fig. 1a, upper right) demonstrated sub-micromolar binding affinity for PKMYT1. The dimethyl phenol was found to be crucial for activity, as evidenced by the complete loss in potency when replaced with a 2,4-dimethyl-3-pyridinyl group (Supplementary Fig. 2b). Compound 1 was recognized as a new chemotype of hit compound for further structure-activity relationship (SAR) exploration, despite relatively low activity. Structure-based design and DMPK optimization yield an optimized PKMYT1 inhibitor Removal of methylpiperazine in the solvent area and bioisosteric replacement of the core ring resulted in compound 2, with improved PKMYT1 IC50 of 19 nM (Fig. 1a, bottom right). However, it failed to suppress phosphorylation of CDK1 at Thr14 in HCC1569 breast cancer cells characterized by CCNE1 amplification, with IC50 > 10 µM. Further physicochemical profiling indicated that low solubility might account for the diminished cellular activity (Supplementary Fig. 2d). Given the distinct structure of compound 2 compared to previously reported PKMYT1 inhibitors, we determined its X-ray co-crystal structure in complex with PKMYT1 at 2.4 Å resolution (Fig. 1b, Supplementary Table 1). The structure revealed that the pyrimidine nitrogen forms a hydrogen bond with Gln196 in the solvent region. The novel scaffold pyrrolo[2,3-b]pyridine forms two hydrogen bonds with Cys190 in the hinge region, but it lacks the water-mediated hydrogen bond between the core and PKMYT1 conserved among PKMYT1-binding small molecules. A cyano group added to the pyridine forms a hydrogen bond with the conserved water molecule, creates steric hindrance preventing free rotation, and acts as an electron-withdrawing group to enhance hinge binding, resulting in compound 3 (Fig. 1c). Other substituents, such as carboxamide, NH2, CF3, and Cl, were found to be inferior to the cyano group in terms of both enzymatic activity and cellular potency (Supplementary Fig. 2c). After chiral resolution, the enantiopure compound 3 demonstrated higher permeability (apical-to-basal flux of 5.46 × 10−6 cm/s) and low efflux ratio (ER 1.28), with an improved recovery rate (48.1%) and aqueous solubility (25.8 μM) (Supplementary Fig. 1d). Notably, a 10-fold improvement in PKMYT1 affinity led to significantly enhanced cellular potency, with a pCDK1 IC50 value of 96 nM. However, high clearance (63.6 mL/min/kg, Fig. 1a) and low oral bioavailability (6%) in rats necessitated further modification. ADMET prediction tools anticipated that the pyrimidinyl methyl group would readily undergo oxidative metabolism (Supplementary Fig. 2e). Compound 4 was the result of replacing the methyl group in the solvent region with an amino group, leading to favorable permeability, aqueous solubility (Supplementary Fig. 2d), atropisomer rotational stability (Supplementary Fig. 2f)31 and metabolic stability (CLp 17.7 mL/min/kg, Fig. 1a), and oral bioavailability (51.9%) in rats, as well as excellent kinase selectivity in the 468 kinome panel (Fig. 1a, bottom left, Supplementary Tables 2, 3). Additionally, the tailed primary amino group formed one more potential hydrogen bond with Gln196 (Fig. 1d) and provided a readily accessible tethering site for incorporation into PROTAC molecules. Design and discovery of PROTAC D1 We proceeded with PROTAC designs targeting CRBN as the E3 ligase adapter due to its widespread expression and the proven therapeutic efficacy of clinical PROTACs that recruit CRBN21,32. To generate a virtual 3D ternary complex model of PKMYT1-PROTAC-CRBN, protein-protein docking was conducted using the “Method 4” within the Molecular Operating Environment (MOE) software package33,34. This process involved docking the binary complex structure of PKMYT1-binder (compound 4) with the binary complex structure of CRBN/CRBN-binder (lenalidomide). Before the docking simulation, the output was set to include 100 conformations of the complex structure for thorough analysis. Based on the results, we selected the poses that demonstrated an appropriately close interaction between PKMYT1 and CRBN. Additionally, the 3D cryo-EM complex structure, comprising CRBN, DDB1, CUL4, and RBX1, was superimposed onto our newly modeled ternary complex, using CRBN as the reference point for alignment. The lysine on the surface of PKMYT1 was expected to extend to E2 to facilitate ubiquitination. Among the results, the pose of Lys345 outside PKMYT1 extended to E2 was selected as the suitable one (Fig. 1e). The distance between the PKMYT1-binder and the CRBN-binder in the docking pose was 8 Å, corresponding to a linker length of approximately 8 to 10 atoms. A linker composed entirely of carbon atoms was used to generate a virtual 3D PROTAC molecule and seed iterative modification. Given that most orally bioavailable PROTACs contain semi-rigid linkers35, we opted not to synthesize PROTACs containing linear carbon linkers. Instead, virtual linear linkers were solely employed as a starting point for subsequent linker generation and design. Two E3-binders, lenalidomide and thalidomide, were individually selected and utilized in subsequent linker design. The E3-binder and the PKMYT1-binder were anchored, and ligand structure-based design, driven by the Chemistry42 generative AI platform, was employed to generate linker structures. After analysis of the generated linker fragments, we chose the piperidinylmethenylpiperazinyl linker therein to form compound D1, which was first selected for synthesis and assessment (Fig. 1f). Compound D1 exhibited highly potent binding affinity for PKMYT1 (IC50 4.9 nM) and CRBN (0.37 μM) and robust degradation of PKMYT1 in HCC1569 cells (DC50 1.0 nM) (Supplementary Fig. 3a). Degradation of PKMYT1 was rapid (Supplementary Fig. 3b, 3d) and notably faster than its natural decay rate (Supplementary Fig. 3c, 3d). Compound D1 suppressed pCDK1 with IC50 of 29 nM, which was nearly three times more potent than the inhibitor-only compound 4 (Fig. 1g), underscoring the effectiveness of targeted degradation in modulating signaling pathways. Compound D1 did not degrade GSPT1, IKZF1/3, or CK1α at concentrations up to 1 μM in HCC1569 cells, indicating a high degree of selectivity over known CRBN neomorphic substrates involved in normal tissue development and disease progression (Supplementary Fig. 3e). Although Compound D1 demonstrated moderate clearance (CLp 40.7 mL/min/kg) and acceptable oral bioavailability in mice (67.0% at 60 mg/kg), its poor solubility (FeSSIF <1.56 μM) and high plasma protein binding (PPB Fu 0.6%), necessitated further structural modification. Optimization of drug-like properties results in lead PROTAC D16 We then screened variants of PROTAC D1 incorporating different CRBN binders currently used in clinical-stage PROTACs for in vitro activity, physical properties, and PK profile. Compounds D1-D6 demonstrated similar affinity for PKMYT1, degradation potency, clearance, and inhibitory activity against pCDK1 (Fig. 2a). However, differences were observed in their physical properties, such as solubility (FeSSIF), PPB, oral absorption, and bioavailability, which might be attributable to their varied physicochemical properties, including molecular weight, number of rotatable bonds and topological polar surface area (TPSA)36. SAR exploration of PROTAC. a Screening of CRBN binders. b Screening of linkers. PPB Fu, fraction unbound in mouse plasma protein binding assay. Mouse PK in CD-1 fed female mice, unless otherwise noted, 1 mg/kg in intravenous (IV) group. The wavy line denotes the bond-forming site. CLp, clearance in plasma; PO, oral administration; Vdss, volume of distribution at steady state; AUC0-24h, exposure between 0 to 24 h; DN fAUC, dose normalized free exposure; F, oral bioavailability. Biochemical binding affinity (HTRF assay) and cellular pCDK1 at Thr14 (in-cell western assay) denote IC50 values. Cellular PKMYT1 degradation (Jess assay) denotes DC50 and Dmax values. Means of triplicates are shown with standard deviations omitted for clarity. a IV group. b PO group. c Fasted mice. All the data are presented as a mean of at least two biological replicates; n = 2 for KS, PPB, and n = 3 for PK. Source data are provided as a Source Data file. Removing one carbonyl group from the CRBN binder in degrader D1 yielded compound D2, which exhibited better solubility (FeSSIF 68 μM) and improved clearance in mice (CLp 21.7 mL/min/kg) but poor oral absorption (F 9.1%). The incorporation of a fluorine atom of D3 significantly improved oral exposure and bioavailability over D2. Compound D4, which features a cyclic urea37, did not significantly improve pCDK1 potency and PK. Compound D5, substituting the cyclic urea for a pyrazole, showed favorable PK in mice at low dose but exhibited sublinear exposure at higher oral dose. The aniline analog D6, with the lowest molecular weight and polar surface area (PSA)38, demonstrated comparable cellular potency to D1 and significantly improved solubility of 109 μM FeSSIF, the highest fraction unbound in PPB, and the greatest oral exposure in mice, achieving a superlinear oral bioavailability of 93.9% at 60 mg/kg. D6 exhibited low permeability (0.31 × 10-6 cm/s) and a high efflux ratio of 12.4, suggesting it is likely a substrate for P-glycoprotein (P-gp), which could potentially account for the superlinear increase in exposure observed at the higher dose. A meta-to-ortho shift of the fluorine atom in the CRBN binder of D6 led to a complete loss of degradation activity (compound D7 in Supplementary Table 4), while the introduction of a second fluorine atom in an alternative meta-position elicited a pronounced hook effect at 1 μM (compound D8 in Supplementary Table 4). Other CRBN binders, such as benzamide piperidine-2,6-dione (D9), phenyl dihydrouracil (D10), and tricycle analog (D11), exhibited inferior pCDK1 potency (Supplementary Table 4). This may be due to the formation of dominant ternary complexes with unbalanced affinities, thereby impacting degradation efficiency, assuming similar kinase inhibition. Therefore, compound D6 bearing the highest dose-normalized unbound exposure stood out for further optimization of linker structure to improve PK properties. We then optimized the PROTAC linker domain for length and angular orientation, leveraging the established PKMYT1-CRL4CRBN model for guidance (Fig. 2b). Insertion of cyclobutane into the linker of D6 to generate compound D12 preserved in vitro activity and enhanced volume of distribution in mice (3.7 L/kg), consistent with its increased lipophilicity. Contracting the 4,6-spirocycle to 4,4-spirocycle in compound D13 decreased degradation efficiency in cells, increased clearance, and diminished oral exposure in mice. Increasing the linker length in compound D6 by one methylene group to generate compound D14 restored degradation efficiency and exhibited enhanced pCDK1 inhibitory activity, albeit with decreased PPB free fraction. Altering the piperazine to piperidine in the linker of D6 to augment its basicity, as seen in compound D15, led to improved solubility and an increased volume of distribution (4.2 L/kg). Utilizing a 5,6-spirocycle to rigidify the linker and reduce the number of rotatable bonds, compound D16 showed marked improvement in cellular potency (pCDK1 IC50 18 nM) and physical properties, including solubility and PPB free fraction. In mice, compound D16 displayed an encouraging PK profile, featuring moderate clearance (56.8 mL/min/kg), a substantial volume of distribution (3.8 L/kg), excellent and dose-proportional oral exposure (AUC0-24h 3093/8634 h*ng/mL at 10/60 mg/kg), and a satisfactory oral bioavailability of 53.4/50.0%. The two linkers of D16 and D14 were combined with other CRBN binders, which resulted in PROTACs with inferior cellular potency (Supplementary Fig. 4). We then modeled the structural interaction profiles of PROTAC D16 with PKMYT1 and CRBN proteins (Supplementary Fig. 5a). The CRBN-binder moiety of D16 forms hydrogen bonds with Trp380, His378, and Glu377 of CRBN. The PKMYT1 ligand domain of D16 exhibits similar interactions to ligand 4, forming hydrogen bonds with the conserved water molecule and with the hinge residue Cys190, H-π interactions with Leu116 and Ser193, and π-π interactions with Phe240. Additionally, the protonated nitrogen of the linker forms a hydrogen bond with residue Leu116 in PKMYT1. The effectiveness of PROTACs varies significantly across different cellular and tissue contexts27. D16 demonstrated consistent potency in cells with CCNE1 amplification, including HCC1599 and MDA-MB-157 breast cancer cells, as well as in FBXW7-mutant HuCC-T1 cholangiocarcinoma cells, similar to its effect in HCC1569 cells (Supplementary Table 5). The CRBN-binder of compound D16 demonstrated exceptional chemical stability, with 97% remaining after 24-hour exposure to acidic conditions (pH 1-2), and no racemization was detected by chiral column chromatography. Additionally, compound D16 exhibited satisfactory stability under simulated gastric fluid (SGF) conditions, retaining 87% of its integrity over a 12-hour period, which is considered adequate for oral drug administration. A novel PROTAC D16-M1P2 engages and degrades PKMYT1 Separation via chiral supercritical fluid chromatography (SFC) yielded four diastereoisomers from D16 (Supplementary Fig. 6a). To identify the optimal stereoisomer, we employed a cell-free competitive HTRF binding assay with recombinant PKMYT1 (Fig. 3a). All diastereoisomers demonstrated comparable binding affinities, with IC50 values ranging from 4.1 nM to 5.0 nM (Supplementary Fig. 6b), similar to compound 4 (IC50 of 2.6 nM), the PKMYT1 ligand of D16, and RP-6306 (IC50 of 2.0 nM) (Supplementary Fig. 6c). Minimal differences were observed in both the predicted binding modes of the four isomers within the PKMYT1-PROTAC-CRBN complexes and their measured affinities for PKMYT1 and CRBN (Supplementary Fig. 5b, c). A novel PROTAC D16-M1P2 engages and degrades PKMYT1. a Structure of D16-M1P2 (single diastereomer, linker stereochemistry not defined). b, c Dose-dependent in vitro PKMYT1 degradation and pCDK1 (T14) inhibition after 24 h treatment of four diastereoisomers of D16 in HCC1569 cells visualized (b) and quantified (c) by JESS. Data are representative of 2 independent experiments. d PKMYT1 and pCDK1 (T14) protein levels in HCC1569 cells treated with D16-M1P2 across a time-course as indicated and quantified by JESS. Data are representative of 2 independent experiments. Source data are provided as a Source Data file. All compounds induced dose-dependent reductions of both PKMYT1 and pCDK1 (T14) in HCC1569 cells (Fig. 3b, c). However, diastereoisomer D16-M1P2 demonstrated superior PKMYT1 degradation efficacy with a DC50 of 0.7 nM and a Dmax of 90%, along with the strongest inhibition of pCDK1 (IC50 of 9.0 nM), thus emerging as the optimal lead candidate. We observed a positive correlation between PKMYT1 degradation and pCDK1 (T14) inhibition up to the maximal degradation concentration of D16-M1P2 (Fig. 3c and Supplementary Fig. 6d), agreeing with the established role of PKMYT1 in phosphorylating CDK1 at Thr148. However, maximal PKMYT1 degradation of 80-90% resulted in only 50-80% inhibition of pCDK1 (T14) (Fig. 3c and Supplementary Fig. 6d). This observation suggests that incomplete degradation of PKMYT1 is insufficient for the complete suppression of pCDK1 (T14), consistent with previous findings8. Moreover, as the drug concentration increased beyond the maximal degradation concentration, we observed a negative association, where PKMYT1 degradation decreased (“hook effect”), but inhibition of pCDK1 (T14) continued to increase (Fig. 3c and Supplementary Fig. 6d). These data suggest that at high concentrations where the hook effect occurs, PKMYT1 predominantly forms a binary complex with D16-M1P2, resulting in strong functional kinase inhibition even in the absence of further protein degradation. To further interrogate degradation kinetics, treatment of HCC1569 cells with 100 nM D16-M1P2 caused rapid PKMYT1 degradation, reducing PKMYT1 levels by over 80% within 4 hours and reaching maximal degradation at 24 hours (Fig. 3d), a rate notably faster than its natural half-life of 12.5 hours (Supplementary Fig. 3c, d). In parallel, D16-M1P2 time-dependently inhibited pCDK1 (T14), achieving over 50% inhibition after 4 hours of treatment at 100 nM, which was sustained for at least 24 hours. However, at lower concentrations of D16-M1P2, a discrepancy emerged between PKMYT1 degradation and inhibition of pCDK1 (T14). Specifically, treatment with 1 nM D16-M1P2 for 8 or 24 hours resulted in approximately 60% reduction in PKMYT1 levels, yet this degree of degradation did not translate into significant inhibition of pCDK1 (T14). This suggests that residual PKMYT1 remaining after partial degradation is sufficient to maintain CDK1 phosphorylation, and that effective pathway suppression requires more complete PKMYT1 depletion or additional inhibitory mechanisms. Interestingly, a short (2 hours) treatment with 100 nM D16-M1P2 resulted in a reduction in PKMYT1 comparable to 8-24 hours of 1 nM treatment (~60%) yet produced significantly greater inhibition of pCDK1 (T14) (Fig. 3d). This enhanced effect is likely due to direct inhibition of PKMYT1 kinase activity by D16-M1P2 at higher concentrations, in addition to its role in promoting PKMYT1 degradation. This finding further supports the existence of dual mechanisms—target degradation and direct enzyme inhibition—contributing to pathway suppression at elevated D16-M1P2 levels. Collectively, these results demonstrate that partial PKMYT1 degradation alone is insufficient to fully inhibit pCDK1 (T14). Robust pathway suppression by D16-M1P2 likely requires both substantial PKMYT1 depletion and direct kinase inhibition at higher concentrations. D16-M1P2 demonstrates dual degradation and inhibition We validated the ubiquitin-proteasome system (UPS)-mediated mechanism of PKMYT1 degradation through co-treatment of D16-M1P2 with the proteasome inhibitor MG-132 or the neddylation inhibitor MLN-4924. PKMYT1 degradation was completely abolished in the presence of MG-132 or MLN-4924 (Fig. 4a), as well as by the CRBN- or PKMYT1-binding small molecules, pomalidomide or Compound 4, respectively (Fig. 4b). These findings corroborate the essential role of PKMYT1- and CRBN-dependent neddylation and proteolytic processes in facilitating UPS-mediated PKMYT1 degradation. D16-M1P2 demonstrates a dual MoA of PKMYT1 degradation and inhibition. Modulation of PKMYT1 protein level by D16-M1P2 with and without proteasome inhibitor MG132 and/or neddylation inhibitor MLN-4924 (a) and by D16-M1P2 with and without PKMYT1 inhibitor compound 4 and/or CRBN binder pomalidomide (b) in HCC1569 cells by immunoblotting. Data are presented as the mean (n = 2) for DMSO and D16-M1P2 (0.1 μM) groups. Data are representative of 2 independent experiments. c PKMYT1 and pCDK1 (T14) levels upon washout of D16-M1P2 or RP-6306 (replacing into compound-free media) in HCC1569 cells visualized (top) and quantified (bottom) by JESS. Data are representative of 2 independent experiments. d Dose–response of PKMYT1 enzymatic activity measured by ADP-Glo kinase assay. Data are presented as the mean (n = 2). e Cellular target engagement of PKMYT1 in a NanoBRET assay. Data are presented as the mean (n = 2). f Dose-dependent in vitro PKMYT1 degradation and pCDK1 (T14) inhibition after 24 h treatment. Source data are provided as a Source Data file. Given the catalytic mechanism of action (MoA) and event-driven pharmacology of PROTACs, prolonged inhibition can be achieved without continuous high drug concentrations17. Following 24-hour treatment of HCC1569 cells with 100 nM D16-M1P2 and subsequent washout, PKMYT1 degradation and pCDK1 (T14) inhibition were maintained for at least 24 hours. In contrast, RP-6306 did not significantly induce PKMYT1 degradation, as expected, and its inhibitory effect on pCDK1 (T14) rapidly returned to baseline within 2 hours post-washout (Fig. 4c). This difference reflects that D16-M1P2-mediated target degradation achieves durable pathway inhibition, whereas RP-6306’s reversible binding without degradation leads to a comparatively transient inhibitory effect. We hypothesized that pCDK1 inhibition beyond maximal PKMYT1 degradation was due to direct PKMYT1 inhibition by the POI-binder domain. To delineate the contributions of PKMYT1 degradation and inhibition, we profiled the CRBN-noninteractive analog D16-NMe-M1P2 (Supplementary Fig. 7a) alongside D16-M1P2. Although D16-NMe-M1P2 retained similar PKMYT1 enzymatic inhibition (IC50 of 8.5 nM) and intracellular target engagement (EC50 of 8.0 nM) to D16-M1P2 (IC50 of 7.6 nM, EC50 of 12.0 nM) (Fig. 4d and e) after 24 hours of treatment, it exhibited one-eighth the pCDK1 (T14) inhibitory potency (Fig. 4f and Supplementary Fig. 7b). In parallel, compared to PKMYT1 inhibitors 4 and RP-6306, D16-M1P2 demonstrated weaker biochemical PKMYT1 enzymatic inhibition and intracellular target engagement (Fig. 4d and e, Supplementary Fig. 7c and 7 d), but more potent pCDK1 (T14) inhibition in HCC1569 cells (Fig. 4f and Supplementary Fig. 7e). These findings suggest that PKMYT1 degradation is the predominant and more effective mechanism underlying D16-M1P2’s long-term activity against PKMYT1. However, at higher concentrations (e.g., 1000 nM) where the hook effect of PKMYT1 degradation occurs, D16-M1P2 enhanced the inhibition of pCDK1 (T14) to a level comparable to that of D16-NMe-M1P2 (Figs. 3c and 4f, Supplementary Fig. 7b). This indicates that PKMYT1 inhibitory activity of D16-M1P2 reinforces pCDK1 (T14) inhibition when PKMYT1 degradation is inefficient, explaining the observed discrepancy between PKMYT1 degradation and pCDK1 (T14) inhibition at these concentrations. D16-M1P2 exhibits high specificity for PKMYT1 Next, we characterized the specificity of D16-M1P2 and found only 4 of the 403 kinases tested, including PKMYT1, exceeded 65% inhibition at 1000 nM (130× its enzymatic IC50) (Fig. 5a, Supplementary Table 2). In contrast, PKMYT1 ligand 4 and RP-6306 were found to bind 11 and 46 off-target kinases, respectively (Fig. 1a, Supplementary Fig. 8a, Supplementary Table 2), demonstrating the significantly improved selectivity of the PROTAC. Further validation of the 3 off-targets, BRAF, RAF1, and SRC, revealed that D16-M1P2 exhibited over 50-fold selectivity for PKMYT1 (IC50 7.6 nM, Fig. 4d) over these kinases (IC50 485.3, 373.1, and 398.5 nM, respectively) (Fig. 5b). Notably, RP-6306 exhibited less than 10-fold selectivity for PKMYT1 (IC50 1.5 nM) over BRAF, RAF1, and SRC (IC50 15.3, 4.4, and 6.4 nM, respectively; Supplementary Fig. 8b), which may contribute to its dose-limiting toxicity in clinical testing. These findings indicate the superior kinase selectivity of D16-M1P2. D16-M1P2 exhibits admirable selectivity. a KinomeScan profile of D16-M1P2 at 1000 nM. Red circles in the TREEspot map (left) indicate kinases found to bind with D16-M1P2 (>65% inhibition), listed in the table (right). b Dose–response of BRAF, RAF1, and SRC enzymatic activity to D16-M1P2 as measured with the ADP-Glo kinase assay. Data are presented as the mean (n = 2). c Change in protein abundance in HCC1569 cells treated with 10 nM or 100 nM D16-M1P2 (n = 3) compared to DMSO-treated controls. Statistical analysis was conducted by a two-sided, unpaired Student’s t-test between D16-M1P2 and DMSO treatments, using p < 0.001 for a stringent significance threshold. Source data are provided as a Source Data file. Because PROTAC molecules may degrade off-target proteins with moderate affinity (≥1–500 nM)17, we profiled the HCC1569 proteome following 8-hour treatment with 10 and 100 nM D16-M1P2. Our analysis quantified 132,593 unique peptides representing 8,177 proteins. PKMYT1 was significantly decreased at both 10 nM and 100 nM treatments (Fig. 5c), which was further validated by JESS analysis (Supplementary Fig. 8c). The majority of other differential proteins (P value < 0.001, log2 FC ≤ −1 or ≥1) were found to exhibit discordant changes in their top two or three most abundant precursor ions, most of which exhibited only minor changes following D16-M1P2 treatment, or an absence of detection within specific groups, indicative of false-positive identification due to dropout (Fig. 5c; Supplementary Fig. 8d), again demonstrating that D16-M1P2 is highly specific to PKMYT1. D16-M1P2 inhibits the growth of tumors with CCNE1 amplification or deleterious alterations in FBXW7/PPP2R1A We conducted cell growth inhibition assays in a panel of 11 tumor cell lines bearing CCNE1 amplification, deleterious alterations in FBXW7 or PPP2R1A, or none of these (Supplementary Fig. 9a), treated with D16-M1P2 or RP-6306. D16-M1P2 displayed, on average, a 12-fold greater selectivity for cells with CCNE1 amplification, FBXW7 mutation, or PPP2R1A loss compared to wild-type cells (IC50 177 nM vs. 2121 nM). In comparison, RP-6306 exhibited a similar potency against biomarker-positive cells (IC50 186 nM) but a relatively lower 6-fold selectivity for cells with these genetic backgrounds (IC50 186 nM vs. 1027 nM for wild-type cells) (Fig. 6a). Notably, RP-6306 showed significantly greater anti-tumor activity than compound 4 in both biomarker-positive and biomarker-negative cell lines (IC50 503 nM vs. 2877 nM for compound 4) (Fig. 6a, Supplementary Fig. 9b), despite displaying similar PKMYT1 enzymatic inhibitory potency (IC50 1.5 nM for RP-6306 vs. 2.0 nM for compound 4). In conjunction with KinomeScan profiling, which revealed lower kinase selectivity for RP-6306 (Supplementary Table 2), these results suggest that its increased potency may be partially attributable to off-target effects. Nevertheless, compound 4 did not demonstrate improved cellular selectivity relative to RP-6306 (both exhibiting a 6-fold selectivity) (Fig. 6a), likely due to its diminished overall potency, which limited the selectivity window. In contrast, D16-M1P2 demonstrated more potent anti-tumor activity than compound 4 in biomarker-positive cells, but not in biomarker-negative cells (Supplementary Fig. 9b), despite having weaker enzymatic potency (Fig. 4d, Supplementary Fig. 7c), illustrating that PKMYT1 degradation can compensate for lower intrinsic inhibitory activity and is the dominant antiproliferative mechanism of D16-M1P2. D16-M1P2 inhibits the growth of tumors with CCNE1 amplification or deleterious alterations in FBXW7 or PPP2R1A and PK profiling. a IC50 values for tumor growth inhibition across a panel of cancer cell lines with or without CCNE1 amplification or deleterious alterations in FBXW7 or PPP2R1A treated with D16-M1P2, RP-6306 or Compound 4 for 7 days. Each cell line is labeled by a dot and the corresponding IC50 value. Data are presented as mean ± SEM. For D16-M1P2 and RP-6306, n = 7 (aberrant) and n = 5 (normal) individual cell lines; for compound 4, n = 5 (aberrant) and n = 3 (normal) individual cell lines. Statistical significance was assessed by a two-side, unpaired Student’s t-test. b In vivo growth inhibition of twice daily oral D16-M1P2 or RP-6306 in HCC1569 xenografts in NOD/SCID mice. Data are presented as mean ± SEM (n = 6 individual mice per group). Statistical significance at endpoint was determined by a two-tailed, unpaired Student’s t-test versus the vehicle group. c–e In vivo PKMYT1 degradation and pCDK1 (T14) inhibition in HCC1569 xenografts treated with either vehicle or twice daily oral D16-M1P2 or RP-6306. Tumor tissues were harvested at 8 or 24 h post the last dosing. The PKMYT1 degradation and pCDK1 (T14) inhibition were visualized (c) and quantified (d, e) by JESS. In d, the correlation between PKMYT1 degradation and pCDK1 (T14) inhibition was evaluated using Pearson’s correlation coefficient, with a two-tailed p value reported. In e, data are presented as mean ± SEM (n = 3 individual mice per group at each timepoint). Statistical differences between groups were evaluated using a two-tailed, unpaired Student’s t-test. f PK profiles for compounds D6 and D16-M1P2 in mice (n = 3), rats (n = 3), dogs (n = 2) and monkeys (n = 2). Same legend as Fig. 2. Source data are provided as a Source Data file. We further evaluated the in vivo efficacy of D16-M1P2 compared to RP-6306. NOD/SCID mice with subcutaneous CCNE1-amplified HCC1569 tumors were randomized to receive D16-M1P2, RP-6306 or vehicle orally. D16-M1P2 exhibited dose-dependent tumor growth inhibition, achieving 35.1% and 66.4% inhibition at 40 mg/kg and 120 mg/kg dosed twice daily for 21 days, respectively (Fig. 6b). In comparison, RP-6306 achieved 71.1% tumor growth inhibition at 20 mg/kg, consistent with previously reported data8. Pharmacokinetic and pharmacodynamic analyses showed that D16-M1P2 induced dose-dependent PKMYT1 degradation and corresponding pCDK1 (T14) inhibition with favorable plasma exposure (Fig. 6c–e, Supplementary Fig. 9c). Unexpectedly, at doses yielding comparable pCDK1 (T14) inhibition (40 mg/kg for D16-M1P2 vs. 20 mg/kg for RP-6306), D16-M1P2 resulted in weaker tumor growth inhibition; similar efficacy was only achieved at a higher dose (120 mg/kg), which provided deeper and more sustained pCDK1 (T14) inhibition (Fig. 6b–e). The inability of D16-M1P2’s pharmacodynamic advantages to translate into superior efficacy is likely attributable to the off-target activity of RP-6306 that may contribute to its antitumor effects beyond reduction in pCDK1. In line with RP-6306’s lower selectivity and therefore increased potential for safety concerns, it caused greater body weight loss at doses required for efficacy comparable to D16-M1P2 (Supplementary Fig. 9d), indicating questionable tolerability and a narrower therapeutic window compared to D16-M1P2. D16-M1P2 exhibits favorable PK profiles across preclinical species To evaluate the drug-like properties of D16-M1P2, we measured PK parameters across preclinical species. Upon intravenous (IV) administration, D16-M1P2 showed low clearance rates across rats, dogs, and monkeys (Fig. 6f). Oral dosing of D16-M1P2 resulted in favorable exposures, achieving acceptable to high bioavailability. Compared to D6 with a moderately rigid linker, D16-M1P2 bearing a rigid spirocyclic linker exhibited superior oral absorption and bioavailability. This underscores the significance of reducing a rotatable bond and introducing an ether ring in the PROTAC linker for enhancing biophysical performance. Discussion Our findings highlight the potential of utilizing AI-enabled platforms for both the discovery of novel kinase inhibitors and PROTAC design. Compared to reported selective PKMYT1 inhibitors, which all feature inner- or outer-cyclic amides as the hinge-domain binder, compound 4 displayed a completely distinct nitrogen heterocycle as the hinge-domain binder. The further development into the bifunctional degrader D16-M1P2 represents a promising advance in balancing potent inhibition with extreme specificity, signifying its potential as a pure chemical probe for the scientific community to explore PKMYT1 biology. The rapid reduction of PKMYT1 levels by 90% at low nanomolar concentrations lasting beyond 24 hours post-washout underscores its potential as a chemical knockdown strategy. D16-M1P2 functions as a dual degrader and inhibitor, with degradation being the predominant and more effective driver. This PROTAC exhibited stronger cellular potency and sustained PD effects compared to the POI-binder alone. D16-M1P2 inhibited proliferation across cancer cell lines with favorable selectivity for CCNE1-amplified, FBXW7-, and PPP2R1A-defective cells, indicative of synthetic lethality between PKMYT1 inhibition and dysregulated cell cycle regulators. D16-M1P2 exhibited antitumor efficacy in xenograft mouse models, with favorable PK properties and prolonged pCDK1 suppression. Moreover, the observed PK properties across four preclinical species, including reasonable clearance rates, favorable volume of distribution, and moderate to high oral bioavailability, further enhances its potential for clinical application. A limitation of our study lies in the observation that, despite the clear mechanistic advantages of the degrader D16-M1P2, manifested as more potent and durable downstream signaling suppression at comparable or lower target occupancy, along with improved selectivity and tolerability relative to RP-6306, the overall antiproliferative potency in vitro and antitumor efficacy in vivo remain similar between these agents under the tested conditions. This apparent disconnect is likely attributable to the PROTAC “hook effect,” which constrains maximal target degradation at higher compound concentrations, thereby limiting the full therapeutic potential of D16-M1P2. Additionally, possible off-target activities associated with RP-6306 may contribute to its antitumor effects, effectively narrowing the observed difference in efficacy. However, inhibition of multiple kinases, including RP-6306’s non-PKMYT1 targets such as BRAF and SRC, is known to cause complex, often dose-limiting adverse effects due to synergistic or additive disruption of overlapping signaling pathways39–41. These toxicities have long limited the clinical utility of multi-kinase inhibitors, despite robust efficacy observed in preclinical models. As such, the balance of efficacy and toxicity of current-generation PKMYT1 inhibitors provides opportunity for safer and potentially more effective target engagement strategies. Notably, our data demonstrate that D16-M1P2 possesses an increased kinase selectivity, which may help mitigate unanticipated side effects and improve safety margins. Nevertheless, the observed hook effect remains a significant challenge for further development. These observations collectively highlight the need for continued optimization of both selectivity and degradation properties in future degrader development. Moreover, rational combination strategies may offer a promising approach to further enhance antitumor efficacy while maintaining tolerability and merit deeper investigation in future studies. Overall, our research provides a foundational basis for the continued exploration of PKMYT1 as a drug target and highlights the merits of AI-driven approaches in discovering new chemical structures and accelerating drug discovery processes. The approaches of this study could have broader applications, potentially offering insights into a variety of other targets as well. Methods Synthesis of compounds The synthesis of compounds 1-4 and D1-D16 is laid out in detail in Supplementary Information. Herein, we take the representative preparation of compound D16-M1P2 to exemplify synthetic chemistry. The synthesis began with commercially available compound S13. Protection of S13 followed by Suzuki coupling with S15 yielded the intermediate chloride compound S16. A palladium-catalyzed reaction converted chloride to a cyano-containing compound S17. Treatment with reagent S21 introduced the bromide, and chiral SFC resolution produced enantiopure S23. Subsequent Suzuki coupling with boric acid ester S26, BBr3-mediated deprotection, and simultaneously converting the pyrimidine chloride to bromide afforded key intermediate S28. In parallel, nucleophilic substitution between S43 and S44, followed by chiral SFC resolution, afforded compound S45-M1. Boc deprotection and reductive amination with spirocyclic ketone created the chiral linker, which SFC resolution gave enantiopure S67-M1P2. Final deprotection and individual coupling reactions with S28 produced target compound D16-M1P2. Despite extensive efforts, crystallization attempts for D16-M1P2, including its ternary complex, small molecule form of itself and the intermediate S67-M1P2, were unsuccessful. The binding affinity of warhead S66-M1 with CRBN (Supplementary Fig. 5d) and literature precedent32,38 support assignment of the S-configuration at the CRBN-binding moiety (Fig. 3a). The currently undetermined stereochemistry of the linker does not affect the clarity or readability of the manuscript. Crystallization and data analysis Recombinant PKMYT1 (residues 75 to 362) protein fused to a TEV cleavable N-terminal 6XHis tag was expressed in Escherichia coli and purified as previously described42. The protein sample was mixed with a crystallization solution composed of 0.1 LiCl, 0.1 M HEPES pH 7.5, 20% v/v PEG 400 using the sitting drop vapor diffusion method. The crystal appeared after 2 days at 18 °C. The crystal with the proper size was harvested and fast-cooled in liquid nitrogen for data collection. The crystals producing high-quality diffraction spots were chosen for full dataset collection at the Diamond Light Source. The diffraction data were indexed and integrated with XDS and scaled by aimless. The space group is H32 and the resolution is 2.38 Å. Structure was solved by MR with Phaser. After multiple rounds of refinement with COOT and refmac5, the final coordinate has Rwork/Rfree values as 0.2402/0.2899. The crystallographic parameters and the statistics of data collection and refinement are shown in Supplementary Table 1. The electron density with and omitting the inhibitor are shown in Supplementary Fig. 10. Molecular dynamics (MD) simulation Molecular dynamics (MD) simulation in this study was conducted using GROMACS 2023.3 software43 to investigate the interactions between RP-6306 and PKMYT1. The crystal structure of the PKMYT1-RP-6306 complex (PDB ID: 8D6E)44 was refined using the Structure Preparation module within MOE software. Structural refinements included the addition of missing hydrogen atoms, side chains of residues, and loops, followed by their correction and optimization. Protonation states of the atoms were determined, while the protein termini were capped to ensure stability. The entire structure was subsequently energy-minimized. The 3D structure of RP-6306 was extracted from the refined complex. Structure optimization of RP-6306 was performed using density functional theory (DFT) at the B3LYP-D3/6-311G** level with the Gaussian 16 quantum chemistry package. Restrained electrostatic potential (RESP) charges for RP-6306 were then computed based on the optimized output file using Multiwfn software (version: 3.8(dev))45,46. The force field parameters for RP-6306 were derived using the General Amber Force Field (GAFF), and RESP charges were mapped onto the prepared 3D structure utilizing the Sobtop package (version 1.0(dev))47, which generated the topology file and restraint file for the small molecules required for MD simulations. The amber99SB force field was utilized to generate the parameter files for the protein PKMYT1 within GROMACS. The parameter files for RP-6306 were incorporated into the topology file of the entire model. The entire model was then subjected to periodic boundary conditions (PBC) and solvated in a cubic water box using the TIP3P water model, ensuring a minimum distance of 10 Å between any protein atom and the edge of the box. Counterions were subsequently added to neutralize system charges and achieve a physiological salt concentration of 0.15 M. The system was first energy-minimized using the steepest descent algorithm until convergence was reached. Equilibration with atom restriction was then performed in two stages: a 1-ns MD simulation under constant volume and temperature (NVT ensemble) using the v-rescale temperature coupling method, followed by a 1-ns MD simulation under constant pressure and temperature (NPT ensemble) using the Parrinello-Rahman barostat. Finally, the fully equilibrated system without any restriction underwent a 200-ns production MD simulation with a time step of 2 fs. Trajectory coordinates were recorded every 10 ps throughout the simulation. The LINCS algorithm was employed to constrain all bonds involving hydrogen atoms, ensuring the stability of the system during simulations. The trajectory was initially preprocessed to correct for periodic boundary effects using the trjconv tool in GROMACS. Subsequently, the interaction statistics between the protein and ligand were analyzed using the ProLIF package48, which facilitated a detailed assessment of molecular interactions. Novel PKMYT1 inhibitor generation via Chemistry42 platform Pharmacophores of RP-6306 and dasatinib were extracted and combined to create a new pharmacophore model. The structure of PKMYT1 with RP-6306 served as the initial starting point for AI generation via the Chemistry42 platform. The 2,4-dimethyl-3-phenol group of RP-6306 in the back pocket was anchored with each atom precisely assigned specific types. The newly combined pharmacophores in the hinge region and the solvent area were used as the input model for AI-driven generation. During this process, over 40 AI-driven generative models were utilized (Supplementary Fig. 1f). During the generation phase, the commercially available Chemistry42 platform leverages advanced models, including deep learning for recognizing patterns within complex chemical data, reinforcement learning to optimize molecular structures based on predefined properties, and generative adversarial networks to design novel molecules30. To further enhance the pipeline, the platform incorporates a range of built-in filters independent of PLI scores, such as medicinal chemistry filters (MCF), similarity filters, drug-likeness evaluations, synthetic accessibility (ReRSA), and diversity filters (Supplementary Fig. 1f). Additionally, rigorous assessment is applied to generated molecular structures, which are refined through 2D scoring modules covering drug-likeness, metabolic site prediction, and flexibility, and 3D reward modules, including pharmacophore prefilters, anchor point prefilters, and shape similarity prefilters. This comprehensive approach ensures alignment with predefined criteria, optimizing molecular properties for research and development applications. This process led to the generation of a total of 2,023 molecules, which were clustered using the FragFp similarity descriptor embedded in DataWarrior software49. Subsequently, rigid docking simulations were carried out using MOE software, followed by visual inspection to assess their protein-ligand interactions. In parallel, these molecules were fragmented into substructure fragments for statistical analysis using the DataWarrior software. Several fragments corresponding to specific functional groups exhibiting favorable protein-ligand interactions were identified. These functional-group fragments were then combined with a fixed 2,4-dimethyl-3-phenol core, anchored from the structure of RP-6306, resulting in the design of novel compounds. Further docking simulations and visual inspections were performed to evaluate their interactions. Ultimately, one molecule (compound 1) was selected for synthesis based on its predicted interactions and overall potential. ADMET properties of compound designs were predicted using Simulations Plus ADMET Predictor (v10). Docking simulation study of small molecules The compounds docking into PKMYT1 in this paper were performed using the Dock module within the Molecular Operating Environment (MOE) software (version 2022). The co-crystal structure of compound 2 in complex with PKMYT1 was utilized for the study. The preparation proceeded in two steps. First, the complex underwent a rapid preliminary refinement using MOE’s QuickPrep panel: Protonate3D was used for protonation with Asn/Gln/His side-chain flips enabled; hydrogens near the binding site were left unfixed; and energy minimization was performed with the default QuickPrep restraints. In the second step, MOE’s Structure Preparation module was applied to further refine the complex, correcting terminal residues, adjusting hydrogen counts, resolving library mapping issues, and standardizing charges. The 3D conformations of the ligands were generated using the conformational search module in MOE. Subsequently, docking simulations of the ligands with PKMYT1 were carried out using the Dock module within the software. For each ligand, 10 poses were generated as output. After thorough visual inspection, one optimal pose for each ligand was selected and proposed. Protein-Protein docking to achieve a linker-compatible conformation of PKMYT1-CRBN interaction Compound 4 was used as the binder of PKMYT1. The complex structure of CRBN with its binder extracted from PDB 4TZ4 was selected. The protein–protein docking was conducted using the “Method 4” within the MOE package33,34. This process involved docking the binary complex structure of PKMYT1-binder (compound 4) with the binary complex structure of CRBN/CRBN-binder (lenalidomide). Both the binary complex structures were loaded as input, and the SMILES of two binders were added as input. Before the docking simulation, the output was set to include 100 conformations of the complex structure for thorough analysis. For each docking pose, the distance between the binders was analyzed. And the lysine site was checked by the modeling of the proteasome complex. The pose that the binders to PKMYT1 and CRBN are close enough in proximity to be connected by a linker was taken as the best protein–protein pose for PROTAC design. Linker generation for PROTAC design via Chemistry42 platform The distance between the binders in the selected protein–protein docked pose is the key criterion for linker selection. The proper conformation with the closest distance of 8 Å was selected. To design PROTAC, a virtual linker composed entirely of carbon atoms was used to generate a virtual 3D PROTAC molecule. The structures of the E3 binder and the PKMYT1 binder were anchored with specific atom types defined. Ligand structure-based design, driven by artificial intelligence, was employed to generate the linker. An internal fragment database was filtered and established as the privileged fragments dataset. During this process, over 40 AI models were utilized, with the molecular weight restricted to below 900. Additionally, the generation of tetracyclic and bridged bicyclic structures was allowed. The generative model and filters applied in linker generation closely mirrored those utilized in the small molecule generation. Cell lines and cell culture All cell lines were cultured at 37 °C and 5% CO2. HCC1569 (ATCC, #CRL-2330), HCC1599 (Cobioer, #CBP60415), HuCC-T1 (JCRB, #JCRB0425), MKN-1 (JCRB, #JCRB0252), HCC1806 (ATCC, #CRL-2611) and NCI-H1568 (ATCC, #CRL-5876) cell lines were cultured in RPMI-1640 with 10% FBS and 1% penicillin-streptomycin. NCI-H1648 (ATCC #CRL-5882) cells were cultured in RPMI-1640 with 5% FBS and 1% penicillin-streptomycin. U-251MG (ECACC, #09063001) and RKO (ATCC, #CRL-2577) cell lines were cultured in EMEM with 10% FBS and 1% penicillin-streptomycin. LN-229 (ATCC #CRL-2611) cells were cultured in DMEM with 10% FBS and 1% penicillin-streptomycin. KLE (ATCC, #CRL-1622) cells were cultured in DMEM: F12 (1:1) with 10% FBS and 1% penicillin-streptomycin. MDA-MB-436 (Cobioer, #CBP60385) cells were cultured in DMEM with 10% FBS, 1% ITS-G, 16 μg/mL Glutathione, and 1% penicillin-streptomycin. MDA-MB-157 (Cobioer, #CBP60381) cells were cultured in MEM with 20% FBS, 1% non-essential amino acids, and 1 mM sodium pyruvate. SW1463 (ECACC, #90112713) cells were cultured in L-15 with 10% FBS and 1% penicillin-streptomycin. All cell lines used in this study were tested and confirmed to be free of mycoplasma contamination by the following contract research organizations: WuXi AppTec (China), ICE Bioscience (China), and Pharmaron (China). Antibodies The following antibodies were utilized in this study: primary antibodies included anti-CDK1 (Thermofisher, cat# 33-1800, clone A17, lot# ZD390434, 1:50 for Jess-automated immunoblotting), anti-CDK1 (phospho T14) (Abcam, cat# ab58509, polyclonal, lot# 1036093-1, 1:1000 for WB; lot# 1097932-23, 1:50 for Jess, 1:1000 for in-cell western), anti-casein kinase 1 alpha (Abcam, cat# ab206652, clone EPR19824, lot# 1045172-12, 1:1000 for WB), anti-Ikaros (CST, cat# 14859, clone D6N9Y, lot# 1, 1:1000 for WB), anti-GSPT1 (Merck, cat# HPA052488, polyclonal, lot# 000034950, 1:1000 for WB), anti-Aiolos (CST, cat# 15103, clone D1C1E, lot# 4, 1:1000 for WB), anti-GAPDH (Merck, cat# MAB374, clone 6C5, lot# 3988901, 1:2500 for WB), anti-vinculin (CST, cat# 13901S, clone E1E9V, lot# 9, 1:10000 for WB; 1:500 for Jess) and anti-Myt1 (CST, cat# 4282S, polyclonal, lot# 2, 1:1000 for WB; 1:50 for Jess); secondary antibodies included IRDye® 800CW goat anti-rabbit IgG (H + L) (LI-COR Biosciences, cat# 926-32211, polyclonal, lot# D30627-05, 1:2000 for in-cell western), anti-rabbit detection module (ProteinSimple, cat# DM-001, polyclonal, lot# 81458, undiluted for Jess), anti-mouse detection module (ProteinSimple, cat# DM-002, polyclonal, lot# 21374, undiluted for Jess), goat anti-mouse IgG H&L (HRP) (Abcam, cat# ab205719, polyclonal, lot# 1093066-3, 1:5000 for WB), goat anti-rabbit IgG H&L (HRP) (Abcam, cat# ab205718, polyclonal, lot# 1036958-1, 1:20000 for WB), goat anti-rabbit IgG (HRP) (CST, cat# 7074, polyclonal, lot# 31, 1:5000 for WB) and horse anti-mouse IgG (HRP) (CST, cat# 7076, polyclonal, lot# 38, 1:5000 for WB). PKMYT1 HTRF assay The test compounds were diluted to 100 μM and then 3-fold serially diluted to achieve 10 concentration gradients. 150 nL of the diluted compound solution was added to the 384-well assay plate by Echo, each dose with 2 replicates. Human recombinant PKMYT1-GST kinase (Carna Bioscience, #05-176) was diluted in 1× kinase reaction buffer to 30 nM, and 5 μL was added to the assay plate and incubated at 25 °C for 15 min. Tracer178 (Invitrogen, #PV5593) was diluted in the 1× kinase reaction buffer to 60 nM, and 5 μL was added to the assay plate and incubated at 25 °C for 60 min. GST-Tb (Revvity, #61GSTTLB) was diluted in 1×PPI buffer to 6 nM and 5 μL was added to the assay plate and incubated at 25 °C for 60 min. The final concentrations were 10 nM PKMYT1 (with a purity of 64%), 20 nM Tracer178, and 2 nM GST-Tb. The signal was measured on Envision 2104 multi-label Reader (PerkinElmer) at 665/615 nm. The IC50 was determined for each compound screened. PKMYT1 ADP-Glo assay PKMYT1 enzymatic activity was detected by measuring ATP hydrolysis using a commercially available ADP-Glo assay (Promega, #V9103) with recombinant human PKMYT1 kinase (ICE, #S2405T-H02G). Briefly, 100 nL of compounds (diluted in DMSO) were added to a 384-well OptiPlate (Revvity, #6007299) using the Echo 655 liquid handler. Next, 5 μL of human recombinant PKMYT1 diluted in reaction buffer (50 mM HEPES, 10 mM MgCl2, 1 mM EGTA, 0.01% Brij35, 2 mM DTT, PH = 7.5) was added to the plate and incubated with the compounds at 25 °C for 10 minutes. Subsequently, inactive recombinant human CDK1 (ICE, #S2211F-H01G), the substrate of PKMYT1 kinase, and ATP were diluted in reaction buffer, and 5 μL was added to initiate the reaction. The plate was centrifuged at 170 × g for 1 minute. The final concentrations were 10 nM PKMYT1, 10 nM unactivated CDK1, and 100 μM ATP. After a 60-minute incubation at 25 °C, 5 μL of ADP-Glo working solution was added, and the plate was centrifuged at 170 × g for 1 minute and incubated for 40 minutes at 25 °C in the dark. Following this, 10 μL of detection working solution was added per well. The plate was centrifuged at 170 × g for 1 min, sealed, and then incubated for a further 40 minutes at 25 °C in the dark. Luminescence was measured using a PHERAstar FSX Microplate Reader (BMG). SRC HTRF assay SRC enzymatic activity was detected using a commercially available HTRF KinEASE TK kit (PerkinElmer, #62TK0PEJ) with recombinant human SRC kinase (ICE, #S2303F-H16G). Briefly, 100 nL of compounds (diluted in DMSO) were added to a 384-well plate (Greiner, #784075) using the Echo 655 liquid handler (Labcyte, #655). Next, 5 μL of 2× recombinant SRC diluted in reaction buffer (1× enzyme buffer from HTRF KinEASE TK kit, 5 mM MgCl2, 1 mM DTT) was added to the plate, centrifuged at 170 × g for 1 minutes, and incubated with the compounds at 25 °C for 10 minutes. Subsequently, 5 μL of 2× TK substrate and ATP solution in reaction buffer was added to initiate the reaction. The final concentrations were 0.3 nM SRC, 1 μM TK substrate, and 50 μM ATP. The plate was centrifuged at 170 × g for 1 minute and incubated at 25 °C for 60 minutes. Following this, 10 μL of 2× XL665 and antibody detection reagents prepared in HTRF KinEASE detection buffer were added. The plate was then centrifuged at 170 × g for 1 minute and incubated at 25 °C for 1 hour. Finally, fluorescence signals at 620 nm (Cryptate) and 665 nm (XL665) were measured using a PHERAstar FSX Microplate Reader (BMG), which allowed for the quantification of kinase activity based on the energy transfer from the donor to the acceptor fluorophore. BRAF and RAF1 ADP-Glo assay BRAF and RAF1 enzymatic activities were assessed by measuring ATP hydrolysis using the ADP-Glo Kinase Assay (Promega, #V9103) with recombinant human BRAF (Carna, #09-122) and RAF1 (Carna, #09-125) kinases. Briefly, 100 nL of compounds (diluted in DMSO) were transferred to a 384-well OptiPlateplate (PerkinElmer, #6007290) using the Echo 655 liquid handler (Labcyte, #655). Then, 5 μL of 2× recombinant BRAF or RAF1 kinase, diluted in assay buffer (50 mM HEPES, 10 mM MgCl2, 0.01% Brij-35, 1 mM EGTA, 2 mM DTT, pH 7.5), was added to each well, centrifuged at 170 × g for 1 minute, and incubated at 25 °C for 10 minutes. Subsequently, 5 μL of 2× unactivated MEK1 substrate (Signalchem, #M02-14BG-50) and ATP in reaction buffer were added to initiate the reaction. The final concentrations were 2 nM BRAF, 1.97 nM RAF1, unactivated MEK1 2.5 ng/μL for BRAF and 1 ng/μL for RAF1, ATP 20 μM for BRAF and 10 μM for RAF1. After centrifuging at 170 × g for 1 minute and incubating at 25 °C for 60 minutes, 5 μL of ADP-Glo working solution was added to deplete remaining ATP, followed by centrifugation at 170 × g for 1 minute and incubation at 25 °C for 40 minutes in the dark. Finally, 10 μL of detection working solution was added, the plate was centrifuged at 170 × g for 1 minute, sealed, and incubated at 25 °C for a further 40 minutes in the dark. Luminescence was measured using a PHERAstar FSX Microplate Reader (BMG), which reflected kinase activity based on the measured signal. NanoBRET assay HEK293 cells were cultured in DMEM with 10% FBS at 37 °C, 5% CO2, and 95% relative humidity until 80-90% confluency. Cells were rinsed with PBS, trypsinized with 0.25% Trypsin-EDTA, and neutralized with DMEM containing 10% FBS. For transient transfection, cells were prepared at a density of 2 × 105 cells/mL in Opti-MEM® I Reduced Serum Medium (Invitrogen, #11058-021) with 1% FBS. A DNA mixture was prepared with 9 μg/mL Transfection Carrier DNA (Promega, #E4881) and 1 μg/mL PKMYT1-NanoLuc Fusion Vector DNA (Promega, #NV1871) in Opti-MEM® I Medium. FuGENE® HD Transfection Reagent (Promega, #E2311) was added at 30 μL/mL of DNA mixture and incubated for 20 minutes. The lipid: DNA complex was mixed with cells at a 1:20 ratio and dispensed into 384-well assay plates (40 μL/well) and incubated for 20–30 hours. For treatment, compounds were dissolved in DMSO to create 1 mM stock solutions, subjected to a 3-fold, 10-point serial dilution using an Apricot liquid handler (Apricot, #PPA-384-M(E)), and transferred (40 nL/well) to 384-well plates with a consistent 0.1% DMSO concentration in all wells, followed by incubation for 30 minutes. A 400 μM Cell-Permeable Kinase NanoBRET® TE Tracer K-5 (Promega, #N2482) solution was prepared, and 100 nL was transferred to each well. The plate was incubated at 37 °C, 5% CO2 for 2 hours. A 3× intracellular TE Nano-Glo Substrate/Inhibitor (Promega, #N2161) was prepared in Opti-MEM® I Medium, and 20 μL was added to each well. The plate was incubated at room temperature for 2–3 minutes. The intensity of the acceptor emission (618 nm) and the donor emission (460 nm) was measured using an EnVision plate reader (Perkin-Elmer). pCDK1 (T14) in-cell western assay HCC1569 cells were seeded at a density of 4000 cells per well in 40 µL of medium in a 384-well plate (Corning, #354663) and incubated overnight at 37 °C with 5% CO2 compounds or DMSO (40 nL) were dispensed into each well using the Echo 655 liquid handler (Labcyte), followed by an additional 24 h incubation. Cells were fixed by adding 40 µL of 8% fixative solution (Solarbio, #P1112) per well and incubating at room temperature for 20 min. After fixation, cells were washed twice with PBS (Solarbio, #P1020) and permeabilized with 40 µL of methanol (Psaitong, #M20033-500) per well. Following two additional PBS washes, cells were blocked with 20 µL of Blocking Buffer (LI-COR, #927-70001) at room temperature for 1 h. Subsequently, the Blocking Buffer was removed, and cells were incubated overnight at 4 °C with a 1:1000 dilution of anti-CDK1 (phospho T14) antibody (Abcam, #ab58509) in Blocking Buffer. The next day, cells were washed three times with PBST (Solarbio, #P1033), 5 minutes each, and then incubated with 20 µL of a mixture containing 1:1000 diluted IRDye 800CW Goat anti-Rabbit IgG (H + L) (LI-COR, #926-32211) in Blocking Buffer at room temperature for 1 h in the dark. After three additional PBST washes (5 minutes each), the plate was centrifuged upside-down at 107 × g for 1 minute and scanned using an Odyssey CLx imaging system (LI-COR). Immunoblotting Cells were lysed using RIPA lysis buffer (Sigma, #R0278) supplemented with protease inhibitor cocktail (Roche, #4693132001) at 4 °C for 20 min. The lysate was collected using a cell scraper and centrifuged at 18,407 × g and 4 °C for 15 min. The supernatant was collected and stored at 80 °C for further immunoblotting analysis. The protein concentration of the cell lysate was determined using the PIERCE BCA Protein Assay Kit (Invitrogen, #23225). Following denaturation, 50 µg of whole-cell lysate was loaded onto each lane of 4–12% gradient Bis-Tris gels (Invitrogen, #NP0336BOX) for protein separation by SDS-PAGE. The gels were run for 0.5 h at 80 V before being run at 120 V for another 1 h. When electrophoresis was completed, the gels containing target proteins were transferred onto NC membrane (Invitrogen, #IB23001) using the iBlot 2 Gel Transfer Device for 10 min at 20 V. All membranes were blocked with 5% DifcoTM skim milk (Beyotime, #P0216-300G) in TBST buffer at room temperature for 1 h and then incubated with primary antibodies specific to the indicated targets in TBST buffer containing 5% BSA at 4 °C overnight. After primary antibody incubation, the membranes were washed three times with TBST buffer and subsequently incubated with HRP-conjugated secondary antibodies at room temperature for 1 h. Blot images were captured using the instrument Image Quant LAS-4000. Quantitative analysis was performed using the software Image Studio Lite Ver 5.2. Jess-automated immunoblotting For cell lysate preparation, cells were lysed using cell lysis buffer (CST, #9803) supplemented with 1× Protease Inhibitor Cocktail (Thermofisher, #1861279) and 1× Phosphatase Inhibitor Cocktail (Thermofisher, #78427) on ice for 30 min, with gentle tapping of the plate every 10 min to ensure thorough lysis. The lysate was collected using a cell scraper and centrifuged at 13,523 × g and 4 °C for 20 min. The supernatant was carefully transferred to a fresh 1.5 mL microcentrifuge tube, aliquoted, and stored at −80 °C until use. For tissue lysate preparation, tumor tissues were transferred into 1.5 mL microcentrifuge tubes, washed twice with ice-cold PBS, and then transferred to a 2 mL microcentrifuge tube containing lysis buffer (CST, #9803) supplemented with magnetic beads. Lysis buffer (100 μL per 10 mg tissue) was added to the sample tube, followed by mechanical dissociation using a cryogenic grinder (Gallop) at −20 °C to prevent protein degradation. After complete homogenization, samples were incubated on ice for 30 min and centrifuged at 12,000 × g for 20 min at 4 °C. The supernatant was carefully transferred to a fresh 1.5 mL microcentrifuge tube, aliquoted, and stored at −80 °C until use. The protein concentration of each lysate was quantified using the BCA Protein Assay Kit (Invitrogen, #23225). Each sample was diluted to a final concentration of 1 mg/mL using lysis buffer and mixed with 5× Fluorescent Master Mix (prepared with 400 mM DTT; EZ Standard Pack, #PS-ST01EZ-8). The mixture was vortexed, heat-denatured at 95 °C for 5 min, vortexed again, and briefly centrifuged at room temperature. A 12–230 kDa Separation Module (ProteinSimple, #SM-W004) was used for size-based protein separation. 3 µL of each sample was loaded into the Jess system’s capillaries. Protein separation, immobilization, and immunodetection were performed automatically within the JESS instrument (ProteinSimple) following the manufacturer’s instructions. Primary antibodies were diluted 1:50 in Antibody Diluent 2 (ProteinSimple). HRP-conjugated secondary antibodies were applied and incubated for 30 min. All incubation steps, including blocking and washes, were automated and controlled by the Jess system to ensure reproducibility. Chemiluminescent Detection was used for target protein visualization. Data were analyzed using Compass Software (ProteinSimple) for automated band quantification. Relative protein expression was normalized to vinculin. Quantitative mass spectrometry A total of 9 HCC1569 cell pellets (3 biological replicates per group: DMSO control, D16-M1P2 10 nM, and 100 nM) were lysed using lysis buffer containing 8 M urea, 50 mM Tris-HCl (pH 7.4), and 1% Triton X-100, supplemented with 0.5% protease and 1% phosphatase inhibitors (Merck, #539134 and #524625). The lysates were vortexed, transferred to grinding tubes, and homogenized using a Fast-24 grinder (Lab Serv). Cell lysates were centrifuged at 14,000 × g for 5 minutes at 4 °C, and the supernatant was collected. Protein concentrations were determined using a BCA kit (Thermo Fisher Scientific, #23227). For protein reduction, 100 μg of protein was incubated with 10 mM dithiothreitol (DTT) at room temperature for 60 minutes. Following reduction, proteins were alkylated with 40 mM iodoacetamide (IAM) and incubated in the dark at room temperature for 30 minutes. Excess IAM was neutralized with an additional 5 mM DTT, and the mixture was incubated for another 30 minutes at room temperature. Hydrophilic and hydrophobic SP3 beads (Cytiva, #45152105050250 and #65152105050250) were mixed at a 1:1 ratio and added to the alkylated sample with additional absolute ethanol to a final concentration of 50% (v/v). Samples were incubated with gentle shaking at room temperature for 15 minutes to allow protein binding. The SP3 beads with bound proteins were washed three times with 80% ethanol and air-dried. The proteins bound to the beads were resuspended in a digestion buffer containing 50 mM HEPES pH 8.0 (SIGMA, #H3375). Trypsin (Promega, #V511A) was added at a protein: enzyme ratio of 50:1 (w/w), and digestion was carried out overnight at 37 °C with continuous shaking at 1,118 × g. Post-digestion, samples were centrifuged at 14,000 × g for 5 minutes and placed on a magnetic rack to allow the beads to settle. The supernatant was collected, and peptide concentrations were determined using a BCA assay. 5 µg of peptides from each sample were desalted using an MS 96-well SPE plate (biocomma, #MC189603), dried using a vacuum concentrator, and stored at −80 °C. LC-MS/MS analysis All nanoLC-MS/MS experiments were performed on an Orbitrap Exploris 480 (Thermo Scientific) equipped with an UltiMate 3000_UPLC system (Thermo Scientific). Peptides were loaded onto a 75 μm id × 2 cm trap column (Acclaim PepMap 100, Thermo Scientific), then separated onto a 75 μm id × 20 cm C18 column packed with reversed-phase silica (Reprosil-Pur 120 C18-AQ, 1.9 μm, Dr. Maisch GmbH). The peptides bound on the column were eluted with a 120-minute linear gradient. The solvent A consisted of 0.1% formic acid in a water solution, and the solvent B consisted of 80% acetonitrile and 0.1% formic acid. The segmented gradient was 6% B, 1.8 min; 6-6.5% B, 0.2 min; 6.5–28% B, 83 min; 28–38% B, 21 min; 38–90% B, 4 min; 90% B, 5 min; 90–4% B, 0.5 min; 4% B, 4.5 min at a constant flow rate of 400 nL/min. The MS analysis was performed with an Orbitrap Exploris 480 mass spectrometer (Thermo Scientific). With the data-independent acquisition mode, the MS scans were acquired at a high resolution of 120,000 (m/z 200) across the mass range of 400-1100 m/z. The AGC target value was 3.00E + 06 with a maximum injection time of 22 ms. The MS/MS scans were acquired with forty variable overlapping precursor isolation windows and centered at: 410, 427.5, 442.5, 455.5, 467, 478, 488.5, 498, 507.5, 517.5, 527.5, 537, 546.5, 556.5, 566.5, 576, 585.5, 595.5, 606.5, 617.5, 627.5, 638.5, 650, 661, 672, 683.5, 696, 709.5, 724, 740, 757.5, 775.5, 794, 813.5, 837.5, 868.5, 907, 953.5, 1009 and 1089.5. Programmed normalized collision energy was set to 25, 27, and 30%. The AGC target value was 1.00E + 06 with a maximum injection time of 40 ms in MS2. For nano electro-spray ion source setting, the spray voltage was 2.1 kV; no sheath gas flow; the heated capillary temperature was 320 °C. Raw MS data were analyzed using DIA-NN software50 (version 1.8.1) against the Swissprot_human protein database51 (UP000005640_9606.fasta, 20,598 entries, downloaded May 7, 2024). The parameters for data analysis were as follows52,53: trypsin/P was specified as the proteolytic enzyme, with up to two missed cleavage sites allowed. Peptide length was set between 7 and 50 amino acids, and a minimum of 1 unique peptide for protein identification. Carbamidomethylation of cysteine was set as a fixed modification, with oxidation of methionine and N-terminal protein acetylation as variable modifications. The precursor FDR was set to 1%. Protein identification and quantification were performed using the top three most abundant unique precursor ions for each protein. Entries were filtered to include at least 2 unique peptides per protein and three valid values in one condition, and the remaining missing values were imputed from the global minimum value. Statistical analysis was conducted using a Student’s t-test between D16-M1P2 and DMSO treatments, using P < 0.001 for a stringent significance threshold. Cellular proliferation assay Cells were seeded into 96-well plate and incubated at 37 °C, 5% CO2 for overnight. The next day, 1 μL of serially diluted compound (final 0.3% DMSO) was added to the plate by Echo 655 and incubated for 7 days in the incubator. On the 7th day, the plate was removed from the incubator and balanced to room temperature. 60 μL of CellCounting-lite reagent was added to each well and incubated at room temperature for 30 minutes in the dark. The luminescence signal was read using a BMG plate reader. Animal studies NOD.CB17-Prkdcscid/NcrCrl mice (female, 6–8 weeks old, ~23–26 g) were purchased from Vital River (Beijing, China). CD-1 mice (female, 6–8 weeks old, ~21–27 g) were purchased from Zhejiang Vital River Laboratory Animal Technology Co., Ltd. Mice were housed in a special pathogen-free environment and in individual ventilation cages (3 mice/cage) within animal rooms maintained on a 12:12 h, light: dark cycle. Animal rooms were temperature and humidity-controlled, between 20~26 °C and 30~70 %, respectively. Mice received food and water ad libitum and were allowed to acclimate for approximately one week before being used for experiments. Sprague−Dawley rats (male, 6-8 weeks old, ~190–290 g) were purchased from Zhejiang Vital River Laboratory Animal Technology Co., Ltd and housed in specific pathogen-free individually ventilated cages (3 rats/cage). Beagle dogs (male, 2–5 years old, ~6–12 kg) were purchased from Beijing Marshall Bio-technology Co., Ltd and individually housed (1 dog/cage). Cynomolgus monkeys (male, 4–9 years old, ~4–8 kg) were purchased from Guangxi Xiongsen Primate Laboratory Animal Breeding and Development Co., Ltd or Hainan Jingang Biotech Co., Ltd. and individually housed (1 monkey/cage). Animals were acclimatized for 1 week (rodents) or underwent a washout period of at least 7 days (non-rodents) prior to experiments. Mouse, rat, dog, and monkey PK study Female CD-1 mice (n = 3) or male Sprague−Dawley rats (n = 3), male Beagle dogs (n = 2) or male Cynomolgus monkeys (n = 2) received either a single intravenous (bolus) injection or a single oral administration (unfasted for rodents, administered 1 h after a meal for non-rodents) of the compound in a cocktail preparation. Doses of indicated compounds were given as solutions in different formulations as follows: (IV) 5% DMSO + 95% (10% HP-β-CD in Saline), (PO) 5% DMSO + 10% Solutol + 85% saline [except that D16-PO-60 mg/kg in 5%DMSO + 15%PEG400 + 10%Solutol+ 70% 50 mM pH4.5 acetate buffer, and D16-M1P2 in 20%PEG400 + 10%solutol+ 70% (20%SBE-β-CD in 50 mM pH4.5 acetate buffer)]. Consecutive blood samples were taken via a submandibular vein or other suitable veins per route of administration after 0.083 (IV only), 0.25, 0.5, 1, 2, 4, 8, and 24 h and were further processed to obtain plasma. Efficacy study of HCC1569 models 3 × 106 HCC1569 cells per mouse were implanted into the central right flanks of female NOD/SCID mice. Compound D16-M1P2 was dissolved in a vehicle solution consisting of 20% PEG400, 10% solutol, and 70% (20% SBE-β-CD in 50 mM pH 4.5 acetate buffer) and was administered orally. Treatments were initiated when the mean tumor size reached 100–200 mm3. Mice were randomly assigned to separate groups (n = 6 mice per group), ensuring comparable mean tumor volumes across all groups at the start of treatment. Tumor size was measured twice a week using a caliper. Tumor volume (mm3) was calculated using the formula: TV = a × b2/2, where ‘a’ represents the longest tumor diameter and ‘b’ is the shortest. These tumor volumes were utilized to calculate tumor growth inhibition (TGI) as follows: TGI (%) = [1-(Tn-T0)/(Vn-V0)] × 100%, over the dosing period (from Day 0 to Day n of treatment). Here, Tn represents the average tumor volume in the treatment group on a given day, T0 is the average tumor volume at the start of treatment, Vn is the average tumor volume in the vehicle group on the same day as Tn, and V0 is the initial average tumor volume of the vehicle group. The maximal subcutaneous tumor size/burden permitted by the Institutional Animal Care and Use Committee (IACUC) of WuXi AppTec was defined as either a mean tumor volume of 2,000 mm³ in the vehicle control group or a tumor volume exceeding 3,000 mm³ in any individual mouse. In this study, neither of these limits was exceeded. Ethics statement All animal care and experimental procedures were conducted in compliance with ethical guidelines and were reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of Medicilon (protocol ID: 001306), WuXi AppTec (protocol ID: ON01-SH006-2023v1.0) or Suzhou Frontage New Drug Development Co., Ltd. (protocol ID: FORM-IP003C-03), in accordance with the regulations of the Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC). Sex-based analysis was not included in this study. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Source data Source data Acknowledgements We thank Min Zheng, Hongyun Qi, Yangguang Li, and Deheng Sun for their contributions to this study. We also thank the contract research organization, Medicilon (China), Viva Biotech (China), ICE Bioscience (China), Pharmaron (China), WuXi AppTec (China), Vivomass Biotech (China), and Eurofins (USA), for performing the synthesis, crystallization, in vitro, and in vivo experiments. Author contributions A.Z. and F.R. conceptualized and supervised the study. X.D. and X.C. guided the study and data analysis. Y. W., X. W., and T.L. performed the research, analyzed the data, and wrote the paper. C.W., Q.M., F.M., J.Y., and J.L. designed the experiments. Y.F. performed the crystallization study. D.G. and F.P. edited the manuscript and provided academic support. M.Z. supervised biology. A.A. supervised the AI platform. All authors approved the final version of the manuscript. Peer review Peer review information Nature Communications thanks Zoran Rankovic, Matthew Robers, Ernst Schonbrunn, Matthew Warren, and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Data availability All data generated in this study, underlying all reported figures, are provided in the main text and Supplementary Information. The cocrystal structure of compound 2 bound with PKMYT1 data generated in this study has been deposited to the Protein Data Bank under accession code 9M6P. The mass spectrometry proteomics data have been deposited to the ProteomeXchange54 Consortium via the PRIDE55 partner repository with the dataset identifier PXD062889. The remaining data are presented in Supplementary Information or Source Data File. Source data are provided with this paper as a Source Data file. Molecular dynamics simulation trajectories are also available as Supplementary Data 1. Source data are provided with this paper. Code availability The code for benchmarking the generative chemistry models is available in the provided reference56. Chemistry42 is an industry-grade commercial software platform available at http://chemistry42.com. We conducted MD simulations and related analyses with open‑source packages (GROMACS 2023.3, Multiwfn 3.8 [dev], SOBTOP 1.0 [dev], PrlLIF) and licensed proprietary software (MOE 2022; Gaussian 16). No third‑party source code was modified or redistributed. Software licenses and citations are listed in the repository README and cited in the manuscript. Any adapted example or helper scripts retain their original license headers. Competing interests Insilico Medicine is a company developing an AI-based end-to-end integrated pipeline for drug discovery and development that is engaged in drug discovery programs for aging, fibrosis, and oncology. Insilico Medicine has filed patent applications, including the compounds described herein [(1) Ding, X. et al. Membrane-associated tyrosine-and threonine-specific cdc2-inhibitory kinase (PKMYT1) inhibitors and uses thereof. WO2023155870, published online 24 Aug 2023; (2) Ding, X. et al. Bifunctional compounds for targeting PKMYT1 and methods of use. WO2025040132, published online 27 Feb, 2025]. Insilico Medicine owns the whole patent rights. All authors are current employees of Insilico Medicine. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Yazhou Wang, Xiaomin Wang, Tingting Liu. Contributor Information Xin Cai, Email: xin.cai@insilico.ai. Xiao Ding, Email: xiao.ding@insilico.ai. Alex Zhavoronkov, Email: alex@insilico.com. Supplementary information The online version contains supplementary material available at 10.1038/s41467-025-65796-8. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Source data Data Availability Statement All data generated in this study, underlying all reported figures, are provided in the main text and Supplementary Information. The cocrystal structure of compound 2 bound with PKMYT1 data generated in this study has been deposited to the Protein Data Bank under accession code 9M6P. The mass spectrometry proteomics data have been deposited to the ProteomeXchange54 Consortium via the PRIDE55 partner repository with the dataset identifier PXD062889. The remaining data are presented in Supplementary Information or Source Data File. Source data are provided with this paper as a Source Data file. Molecular dynamics simulation trajectories are also available as Supplementary Data 1. Source data are provided with this paper. The code for benchmarking the generative chemistry models is available in the provided reference56. Chemistry42 is an industry-grade commercial software platform available at http://chemistry42.com. We conducted MD simulations and related analyses with open‑source packages (GROMACS 2023.3, Multiwfn 3.8 [dev], SOBTOP 1.0 [dev], PrlLIF) and licensed proprietary software (MOE 2022; Gaussian 16). No third‑party source code was modified or redistributed. Software licenses and citations are listed in the repository README and cited in the manuscript. Any adapted example or helper scripts retain their original license headers.