2026.06.18
End-to-End AI-Driven Design of a First-in-Class PKMYT1 PROTAC Using Chemistry42
Author: Haining Wang, Application Scientist | Chemistry42
haining.wang@insilicomedicine.com
haining.wang@insilicomedicine.com
This white paper focuses on the application of Chemistry42 for PROTAC design. Experimental validation is available in the published paper: Wang, Y., Wang, X., Liu, T. et al. Discovery of a bifunctional PKMYT1-targeting PROTAC empowered by AI-generation. Nat. Commun. 16, 10759 (2025).
https://doi.org/10.1038/s41467-025-65796-8
https://doi.org/10.1038/s41467-025-65796-8
Proteolysis-targeting chimeras (PROTACs) leverage an event-driven pharmacological mechanism to enable intervention against targets that are difficult to address with conventional small-molecule inhibitors. This white paper showcases how ISM2286, a first-in-class bifunctional PKMYT1 PROTAC degrader, was designed using the Chemistry42 platform, which was leveraged at each critical stage of PROTAC development.
The core workflow encompasses a multi-parameter optimization (MPO) and molecular design framework that includes:
The core workflow encompasses a multi-parameter optimization (MPO) and molecular design framework that includes:
- 20+ AI generative models for the design of novel warhead scaffolds generation
- Druglike molecular property prediction through over 900 Medicinal Chemistry Filters (MCFs)
- Binding affinity prediction via Pocket-Ligand Interface (PLI) scoring with LE / LLE reference metrics for comprehensive candidate assessment and prioritization
- Early Kinome off-target screening with Golden Cubes
- Medicinal chemistry-guided refinement
- 140 built-in ADMET prediction models for concurrent pharmacokinetic property optimization
1. Project Background
PKMYT1 has emerged as an attractive therapeutic target due to its role in cell-cycle regulation and its synthetic lethality with CCNE1-amplified tumours. Existing PKMYT1 inhibitors, including RP-6306, have validated the druggability and therapeutic relevance of this target. However, a conventional kinase inhibitor strategy still relies on sustained target occupancy and must overcome common challenges of ATP-site kinase inhibition, including selectivity, resistance, and therapeutic-window limitations.
These considerations motivated us to explore a targeted protein degradation strategy for PKMYT1. A PROTAC approach could convert transient target binding into durable PKMYT1 depletion by recruiting an E3 ligase and inducing ubiquitin-proteasome-mediated degradation. In principle, this mechanism may provide deeper and longer-lasting pathway suppression than reversible inhibition alone. It may also improve functional selectivity by promoting productive ternary complex formation, in which the target protein, PROTAC, and the E3 ligase must form a geometrically compatible complex.
However, successful PKMYT1 PROTAC design depends critically on the quality of the warhead. The warhead must not only bind PKMYT1 with sufficient potency, but also provide an appropriate linker exit vector, maintain cellular permeability after conjugation, and avoid broad off-target kinase engagement. Therefore, instead of directly converting RP-6306 into a degrader, the project aimed to design a new PKMYT1 warhead better suited for PROTAC construction.
This rationale led to three major design requirements: improved kinase selectivity, a linker-accessible binding orientation, and a more PROTAC-compatible physicochemical profile.
For detailed target biology, synthetic lethality mechanisms, and the rationale for druggability, please refer to the published paper.
These considerations motivated us to explore a targeted protein degradation strategy for PKMYT1. A PROTAC approach could convert transient target binding into durable PKMYT1 depletion by recruiting an E3 ligase and inducing ubiquitin-proteasome-mediated degradation. In principle, this mechanism may provide deeper and longer-lasting pathway suppression than reversible inhibition alone. It may also improve functional selectivity by promoting productive ternary complex formation, in which the target protein, PROTAC, and the E3 ligase must form a geometrically compatible complex.
However, successful PKMYT1 PROTAC design depends critically on the quality of the warhead. The warhead must not only bind PKMYT1 with sufficient potency, but also provide an appropriate linker exit vector, maintain cellular permeability after conjugation, and avoid broad off-target kinase engagement. Therefore, instead of directly converting RP-6306 into a degrader, the project aimed to design a new PKMYT1 warhead better suited for PROTAC construction.
This rationale led to three major design requirements: improved kinase selectivity, a linker-accessible binding orientation, and a more PROTAC-compatible physicochemical profile.
For detailed target biology, synthetic lethality mechanisms, and the rationale for druggability, please refer to the published paper.
2. Chemistry42 PROTAC Development Workflow
This section provides a detailed description of how the Chemistry42 platform was applied across the three core phases of PROTAC development. The overall workflow is illustrated below:
Figure 1.
Overview of the Chemistry42-driven end-to-end design workflow for the PKMYT1 PROTAC.
2.1
Phase 1: AI-Powered Warhead Generation and Selection
Phase 1: AI-Powered Warhead Generation and Selection
This phase represents the most intensive application of Chemistry42, with multiple core platform capabilities forming a design cycle.
2.1.1
Meticulous Design Input Structure Preparation
Meticulous Design Input Structure Preparation
Directly using RP-6306 as a PROTAC warhead is problematic due to structural homogeneity and resistance risk, poor kinase selectivity, and unfavourable molecular features (a deeply buried binding site and 5 HBDs). The objective of this phase was to leverage Chemistry42 to design a novel warhead with a distinct chemotype that provides improved selectivity, a solvent-exposed linker exit vector, and a compact physicochemical profile suitable for oral PROTAC development. To enable the AI generative models to explore novel scaffolds in the solvent-exposed region freely, we performed the following meticulous input configuration in Chemistry42.
Starting Ligand and Pharmacophore Setup: RP-6306 (8D6E) was used as a reference ligand. Explicit hydrogen-bond pharmacophore constraints were set around its key interactions with PKMYT1 in the hinge-binder region, defining two critical hydrogen bonds (one donor, one acceptor). A mandatory Point was applied to enforce these pharmacophore features as hard constraints, ensuring that all AI-generated scaffolds satisfy these key interactions: a prerequisite for maintaining high affinity to PKMYT1. Additionally, a 2D Privileged Fragment was used to fix the 2,4-dimethyl-3-phenol region, preserving this critical hydrophobic anchor in the back pocket. We deliberately removed the crystallographic water molecules from the region facing the solvent-exposed pocket, enabling Chemistry42 to freely grow fragments in this area and explore novel scaffolds that extend toward the solvent and provide attachment points for linkers.
Starting Ligand and Pharmacophore Setup: RP-6306 (8D6E) was used as a reference ligand. Explicit hydrogen-bond pharmacophore constraints were set around its key interactions with PKMYT1 in the hinge-binder region, defining two critical hydrogen bonds (one donor, one acceptor). A mandatory Point was applied to enforce these pharmacophore features as hard constraints, ensuring that all AI-generated scaffolds satisfy these key interactions: a prerequisite for maintaining high affinity to PKMYT1. Additionally, a 2D Privileged Fragment was used to fix the 2,4-dimethyl-3-phenol region, preserving this critical hydrophobic anchor in the back pocket. We deliberately removed the crystallographic water molecules from the region facing the solvent-exposed pocket, enabling Chemistry42 to freely grow fragments in this area and explore novel scaffolds that extend toward the solvent and provide attachment points for linkers.
Figure 2.
Chemistry42 warhead generation experiment initial setup interface: RP-6306 as input ligand, showing protein structure, pharmacophore constraints, and Mandatory Points configuration
Pocket Dilation (5 Å): To ensure the Chemistry42s have sufficient space to generate new structures in the solvent-exposed region, we applied a 5 Å pocket dilation in Chemistry42. This operation expands the three-dimensional spatial boundary available for AI exploration, enabling the generative models to extend beyond the original binding pocket into solvent-exposed areas, providing physical space for novel scaffolds to expand into.
Figure 3.
Chemistry42 pocket dilation (5 Å) setting interface.
Pharmacophore Configuration: The figure below shows the detailed pharmacophore constraint settings, including hydrogen-bond donor/acceptor definitions.
Figure 4.
Chemistry42 pharmacophore-constraint detailed setup interface, showing precise definitions of the hinge-region hydrogen-bond donors and acceptors.
Dual-Template Pharmacophore Fusion: Building on the above refined configuration, we incorporated the structural features of dasatinib that extend toward the solvent-exposed region (a hydrogen bond with Gln196), constructing a fused pharmacophore model that directs AI generative models to simultaneously explore bifunctional scaffolds that satisfy hinge-region binding and solvent-area extension.
Note: Chemistry42 now supports Protein Pharmacophores, enabling users to set specific protein residues, such as Gln196, as direct pharmacophore constraints. This allows AI-generated models to explore the key interaction space without requiring a reference ligand structure.
Note: Chemistry42 now supports Protein Pharmacophores, enabling users to set specific protein residues, such as Gln196, as direct pharmacophore constraints. This allows AI-generated models to explore the key interaction space without requiring a reference ligand structure.
Key takeaway: Chemistry42’s input configuration goes far beyond simply “loading a reference molecule.” The precise setting of pharmacophore constraints and Mandatory Points, the 2D Privileged Fragment fixing of the phenol region, pocket dilation of 5 Å, expanding the exploration space, strategic removal of structural water molecules, and a multi-template fusion strategy collectively determine Chemistry42’s exploration direction and generation quality.
2.1.2
Iterative Experimental Configuration and Chemistry42 Selection Strategy
Iterative Experimental Configuration and Chemistry42 Selection Strategy
A key practical insight is that during the experimental setup phase, we adopted an iterative strategy in which we tested multiple experimental configuration combinations, adjusting configuration parameters based on the quality of the molecular structures generated in each round and gradually converging on an optimal configuration before conducting the full-scale generation experiment.
Chemistry42 provides three preset AI generative model tiers to accommodate the needs of different stages:
Chemistry42 provides three preset AI generative model tiers to accommodate the needs of different stages:
Figure 5.
Chemistry42’s three preset AI generative model tiers: Light, Optimal, and Advanced.
2.1.3
Multi-Parameter Assessment
Multi-Parameter Assessment
Chemistry42’s built-in multi-parameter scoring system was used to systematically assess the generated molecules:
PLI Score (Pocket-Ligand Interaction Score) evaluates each generated molecule’s theoretical binding energy to the PKMYT1 binding pocket, serving as a first-pass filter to eliminate clearly non-binding compounds.
LE (Ligand Efficiency) and LLE (Lipophilic Ligand Efficiency) Reference Metrics. The thresholds of LE > 0.3 and LLE > 3 are not hard cutoff filters but serve as reference indicators to assist in prioritization.
The reference value of these two metrics reflects the core requirements of an ideal PROTAC warhead from complementary angles: LE reference: A high-quality warhead must not only exhibit strong affinity for the protein of interest (POI), but crucially the molecule must not be overly bulky, the number of heteroatoms should be minimized, and both hydrogen bond donors (HBDs) and acceptors (HBAs) should be kept to a reasonable count. This is because the warhead must ultimately be connected to a linker and E3 binder to form the complete PROTAC molecule; a compact warhead preserves precious molecular weight and polarity budget for the downstream modules. LLE reference: A reasonable LogP value is another critical warhead design consideration. High LLE indicates that molecular activity derives from specific polar interactions rather than non-specific hydrophobic contacts. This provides a wider physicochemical optimization window for subsequent overall PROTAC performance tuning, especially for the development of orally bioavailable PROTAC drugs.
PLI Score (Pocket-Ligand Interaction Score) evaluates each generated molecule’s theoretical binding energy to the PKMYT1 binding pocket, serving as a first-pass filter to eliminate clearly non-binding compounds.
LE (Ligand Efficiency) and LLE (Lipophilic Ligand Efficiency) Reference Metrics. The thresholds of LE > 0.3 and LLE > 3 are not hard cutoff filters but serve as reference indicators to assist in prioritization.
The reference value of these two metrics reflects the core requirements of an ideal PROTAC warhead from complementary angles: LE reference: A high-quality warhead must not only exhibit strong affinity for the protein of interest (POI), but crucially the molecule must not be overly bulky, the number of heteroatoms should be minimized, and both hydrogen bond donors (HBDs) and acceptors (HBAs) should be kept to a reasonable count. This is because the warhead must ultimately be connected to a linker and E3 binder to form the complete PROTAC molecule; a compact warhead preserves precious molecular weight and polarity budget for the downstream modules. LLE reference: A reasonable LogP value is another critical warhead design consideration. High LLE indicates that molecular activity derives from specific polar interactions rather than non-specific hydrophobic contacts. This provides a wider physicochemical optimization window for subsequent overall PROTAC performance tuning, especially for the development of orally bioavailable PROTAC drugs.
Key takeaway: PLI scoring serves as the primary filter, while LE and LLE function as reference indicators for comprehensive assessment. What ultimately determines warhead selection is the ideal profile collectively reflected by these reference metrics: high affinity, a compact molecular scaffold, moderate polarity, and a reasonable lipophilicity profile – ensuring the warhead is “lean and versatile,” leaving ample optimization space for downstream PROTAC development.
ADMET Predictor Initial Screen Chemistry42’s built-in ADMET prediction tools were invoked at this stage to identify metabolic liability sites: for example, the oxidative metabolism hotspot at the pyrimidine methyl group was successfully flagged early, providing advance warning for subsequent optimization.
2.1.4
Golden Cubes Rapid Kinase Screening
Golden Cubes Rapid Kinase Screening
This is the most differentiating capability of Chemistry42 in the context of this project.
Golden Cubes is Chemistry42’s built-in kinase screening module, which rapidly provides virtual predictions of each generated molecule’s inhibitory potential against tens of kinases. The module operates as follows: for each kinase, it outputs a probability value representing the likelihood that the compound inhibits that kinase with an IC50 below 500 nM. A higher probability indicates a greater likelihood of the compound inhibiting that particular kinase. In the visualization interface, kinases with a high probability (i.e., high off-target risk) are marked with green cubes.
Screening criterion in this project: Molecules with 5 or fewer green cubes (i.e., 5 or fewer kinases predicted as likely off-targets) advanced to the next stage.
Golden Cubes is Chemistry42’s built-in kinase screening module, which rapidly provides virtual predictions of each generated molecule’s inhibitory potential against tens of kinases. The module operates as follows: for each kinase, it outputs a probability value representing the likelihood that the compound inhibits that kinase with an IC50 below 500 nM. A higher probability indicates a greater likelihood of the compound inhibiting that particular kinase. In the visualization interface, kinases with a high probability (i.e., high off-target risk) are marked with green cubes.
Screening criterion in this project: Molecules with 5 or fewer green cubes (i.e., 5 or fewer kinases predicted as likely off-targets) advanced to the next stage.
Figure 6.
Chemistry42 Golden Cubes kinase screening module visualization interface. Each cube represents a kinase; green cubes indicate a high probability that the compound will achieve an IC50 < 500 nM for that kinase (indicating a high off-target risk). Screening criterion: 5 or fewer green cubes.
The core value of Golden Cubes lies in compressing what traditionally requires weeks of wetlab experimental screening for kinome selectivity into a matter of hours, enabling early kinome off-target screening. This allows a “selectivity-first” design philosophy to be implemented at the warhead stage, rather than discovering off-target issues only at the final PROTAC molecule stage.
2.1.5
User Manual Refinement (AI + Human Synergy)
User Manual Refinement (AI + Human Synergy)
The AI-generated pyrrolo[2,3-b]pyridine warhead scaffold from Chemistry42 was further refined by medicinal chemists based on structural insights. A cyano (CN) group was introduced at a key position on the pyridine ring, yielding Compound 3. This single modification delivered three benefits: restoring a conserved water-mediated hydrogen bond in the PKMYT1 binding pocket, restricting biaryl rotation to reduce conformational penalty, and enhancing hinge binding through its electron-withdrawing effect. SAR studies showed that alternative groups such as CONH2, NH2, CF3, and Cl were less effective, confirming the unique advantage of CN at this position.
Figure 7.
Warhead Compound 3 was obtained after manual modification involving the introduction of the CN group.
Key takeaway: Chemistry42 enabled efficient exploration of novel chemical scaffolds, which were subsequently refined by medicinal chemists using structure-based insights to optimize binding affinity. The strategic introduction of the CN group underscores the powerful synergy between AI-driven molecular generation and human expertise in medicinal chemistry.
2.1.5
From Compound 3 to Final Warhead ISM4963
From Compound 3 to Final Warhead ISM4963
Although Compound 3 demonstrated excellent PKMYT1 affinity and significantly improved kinase selectivity, its pharmacokinetic profile remained suboptimal (high clearance of 63.6 mL/min/kg, only 6% oral bioavailability in rats). Chemistry42’s ADMET prediction tools identified the pyrimidine methyl group as an oxidative metabolism hotspot.
As such, the terminal carbon (methyl) at the solvent-exposed end of Compound 3 was replaced with an NH2 group, yielding the final warhead: ISM4963
As such, the terminal carbon (methyl) at the solvent-exposed end of Compound 3 was replaced with an NH2 group, yielding the final warhead: ISM4963
Figure 8.
A comparison of ISM4963 and RP-6306. ISM4963 achieves excellent kinase selectivity and pharmacokinetic properties while maintaining high affinity.
Key breakthrough: Unlike RP-6306’s binding mode, which is deep in the pocket, ISM4963’s primary amine tail successfully extends to a position much closer to the solvent-exposed surface, significantly shortening the linker bridging distance. This enables a short ~8 Å linker to bridge to the E3 ligase binder, fundamentally avoiding the flexibility, permeability, and metabolic stability risks associated with excessively long linkers. Additionally, the primary amine provides a natural linker tethering site, offering an ideal chemical attachment point for subsequent PROTAC assembly. The Golden Cubes virtual screening results were subsequently confirmed experimentally, with the number of off-target kinases reduced from 46 to 11 (>76% reduction).
2.2
Phase 2: Ternary Complex Modelling and Spatial Constraint Extraction
Phase 2: Ternary Complex Modelling and Spatial Constraint Extraction
With the warhead established, the next step was to construct a spatial model of the “PKMYT1-PROTAC-E3 ligase” ternary complex to define three-dimensional distance constraints for linker design.
E3 ligase and binder selection: CRBN was selected as the E3 ligase subunit for this project. The E3 binder was selected by medicinal chemists manually based on established structure-activity relationships for the pomalidomide series.
Ternary complex conformational modelling: External computational tools were first used to perform preliminary docking of the PKMYT1-CRBN complex, generating candidate conformations. These candidates were then imported into MD for full-atom molecular dynamics stability assessment, filtering out unstable docking poses, and ultimately selecting an optimal ternary complex model.
E2 subunit introduction and ubiquitination geometry assessment: Following MD stability screening and selection of the ternary complex conformation, the E2 ubiquitin-conjugating enzyme subunit was further introduced using existing experimental structure models or AlphaFold3 predictions to construct a complete PKMYT1-PROTAC-CRL4(CRBN)-E2 complex. This complete model was used to assess whether the target lysine on the PKMYT1 surface (Lys345) is oriented toward E2 in a geometry consistent with productive ubiquitination, a key modelling step for conformational prioritization. It should be noted that this Lys345 ubiquitination model remains a computational hypothesis proposed based on the structural model; experimental validation is still pending. Nevertheless, this structural assessment provides valuable rationale for selecting ternary complex poses most likely to support productive degradation.
E3 ligase and binder selection: CRBN was selected as the E3 ligase subunit for this project. The E3 binder was selected by medicinal chemists manually based on established structure-activity relationships for the pomalidomide series.
Ternary complex conformational modelling: External computational tools were first used to perform preliminary docking of the PKMYT1-CRBN complex, generating candidate conformations. These candidates were then imported into MD for full-atom molecular dynamics stability assessment, filtering out unstable docking poses, and ultimately selecting an optimal ternary complex model.
E2 subunit introduction and ubiquitination geometry assessment: Following MD stability screening and selection of the ternary complex conformation, the E2 ubiquitin-conjugating enzyme subunit was further introduced using existing experimental structure models or AlphaFold3 predictions to construct a complete PKMYT1-PROTAC-CRL4(CRBN)-E2 complex. This complete model was used to assess whether the target lysine on the PKMYT1 surface (Lys345) is oriented toward E2 in a geometry consistent with productive ubiquitination, a key modelling step for conformational prioritization. It should be noted that this Lys345 ubiquitination model remains a computational hypothesis proposed based on the structural model; experimental validation is still pending. Nevertheless, this structural assessment provides valuable rationale for selecting ternary complex poses most likely to support productive degradation.
Figure 9.
Schematic of the PKMYT1-PROTAC-CRL4(CRBN)-E2 complex model. PKMYT1 is shown in green, with Lys345 (green sticks) oriented toward the E2 ubiquitin-conjugating enzyme, satisfying the correct ubiquitination geometry. in a geometry consistent with productive ubiquitination (computational hypothesis; experimental validation pending)
2.3
Phase 3: AI-Designed Linkers
Phase 3: AI-Designed Linkers
With a distance constraint of approximately 8 Å established, the project entered the most challenging phase of PROTAC design and development.
We employed Chemistry42’s Anchor Points to fix the two terminal binders (AI-designed ISM4963 + manually selected CRBN ligand) and deployed a multi-round linker generation strategy. Anchor Points are designated “locked atoms” in Chemistry42’s molecular editor at the two connection endpoints ( see blue spheres in Figure 10). The AI Models, therefore, do not alter the chemical structures at these two ends during generation and explore linker chemical space exclusively between the two Anchor Points.
Similar to the warhead design phase, linker design also required iterative exploration of different platform configurations. We tested multiple configuration schemes, including different constraint combinations, fragment-guidance strategies, and ADMET-optimization weights, progressively optimizing configurations through iterative feedback.
We employed Chemistry42’s Anchor Points to fix the two terminal binders (AI-designed ISM4963 + manually selected CRBN ligand) and deployed a multi-round linker generation strategy. Anchor Points are designated “locked atoms” in Chemistry42’s molecular editor at the two connection endpoints ( see blue spheres in Figure 10). The AI Models, therefore, do not alter the chemical structures at these two ends during generation and explore linker chemical space exclusively between the two Anchor Points.
Similar to the warhead design phase, linker design also required iterative exploration of different platform configurations. We tested multiple configuration schemes, including different constraint combinations, fragment-guidance strategies, and ADMET-optimization weights, progressively optimizing configurations through iterative feedback.
Figure 10.
Chemistry42 linker generation experiment setup interface. Both terminal binders are fixed via Anchor Points (red labels), with AI exploring linker structures only in the intervening region.
2.3.1
Round 1: Unconstrained Chemical Space Exploration
Round 1: Unconstrained Chemical Space Exploration
The core strategy of Round 1 was “unconstrained chemical space exploration”: Chemistry42’s generative models extended from a basic PEG linker starting point, with no additional 2D privileged fragment or substructure constraints beyond Anchor Point fixation and the 8 Å distance constraint. The objective was to comprehensively map which topological connection pathways are physically feasible under the given spatial constraints, acquiring extensive chemical space sampling data.
Key Discovery, Emergence of Piperazine Motif: Among the large number of linkers generated in this unconstrained exploration round, the platform autonomously “discovered” a class of linker structures containing a piperazine ring. Molecular docking analysis revealed that the piperazine nitrogen atoms formed novel hydrogen-bond interactions with protein residues in the linker region. This finding is significant: traditional PEG linkers are typically considered inert connectors, whereas the piperazine-containing linker actively participates in protein binding, providing an additional contribution to ternary complex stability.
Key Discovery, Emergence of Piperazine Motif: Among the large number of linkers generated in this unconstrained exploration round, the platform autonomously “discovered” a class of linker structures containing a piperazine ring. Molecular docking analysis revealed that the piperazine nitrogen atoms formed novel hydrogen-bond interactions with protein residues in the linker region. This finding is significant: traditional PEG linkers are typically considered inert connectors, whereas the piperazine-containing linker actively participates in protein binding, providing an additional contribution to ternary complex stability.
Figure 11.
Schematic showing the novel hydrogen-bond interactions introduced by the piperazine motif in the linker region. The protonated nitrogen of the piperazine ring forms a hydrogen bond with PKMYT1 residue Leu116, actively participating in protein binding.
The initial PROTAC hit from this round, compound D1, demonstrated encouraging cellular activity (DC50 = 1 nM), but its physicochemical properties required further improvement (poor solubility, high plasma protein binding).
2.3.2
Round 2: MCS Privileged Fragment Guidance and Rigidification
Round 2: MCS Privileged Fragment Guidance and Rigidification
Building on the spatial feasibility data from Round 1 (particularly the piperazine hydrogen bond discovery), Round 2 introduced more refined constraints. Multiple different experimental configurations were tested in Chemistry42, with the key configuration that yielded the final PROTAC molecule as follows:
Figure 12.
The privileged Fragments library is used in the Round 2 linker generation. These fragments were prepared by extracting Maximum Common Substructures (MCS) from public PROTAC linker compound libraries and used as 2D constraint conditions to guide AI generation.
MCS Privileged Fragment Strategy in Detail: From publicly available PROTAC linker compound libraries, we extracted Maximum Common Substructures (MCS) representing characteristic linker substructures and curated them as “Privileged Fragments” for import into Chemistry42 as constraint conditions. The core logic of this strategy is that high-frequency substructures in public linker libraries represent experimentally validated chemical patterns with established drug-likeness and synthetic accessibility, and that incorporating them as 2D constraints in the AI generation process guides the models to explore innovative combinations within proven chemical space while maintaining chemical feasibility.
Chemistry42’s built-in ADMET prediction models played a pivotal role at this stage. In conventional PROTAC development, ADMET optimization typically occurs late in the process, leading to high attrition rates due to poor drug-like properties. Chemistry42 front-loads ADMET prediction into the generative process, providing each AI-generated linker with concurrent multi-dimensional property assessments, including solubility, clearance, and permeability, enabling a “generate-and-optimize-simultaneously” workflow.
Breakthrough Modification – Spirocyclic Rigidification: The AI generative models successfully incorporated 5,6-spirocycle rigid structural motifs into the linker, yielding multiple benefits: Rigid structures reduce the number of flexible rotatable bonds, lowering conformational entropy penalties and improving oral absorption - Ether rings provide additional polar surface area, enhancing aqueous solubility - Achieved an optimal balance between rigidity and polarity, overcoming the traditional challenge of poor water solubility in large-molecule drugs.
Ultimately, spirocyclic linker variants achieved a >100-fold improvement in solubility (FeSSIF, Fasted State Simulated Intestinal Fluid) over flexible linkers (reaching 156 μM), with corresponding improvements in oral bioavailability.
Chemistry42’s built-in ADMET prediction models played a pivotal role at this stage. In conventional PROTAC development, ADMET optimization typically occurs late in the process, leading to high attrition rates due to poor drug-like properties. Chemistry42 front-loads ADMET prediction into the generative process, providing each AI-generated linker with concurrent multi-dimensional property assessments, including solubility, clearance, and permeability, enabling a “generate-and-optimize-simultaneously” workflow.
Breakthrough Modification – Spirocyclic Rigidification: The AI generative models successfully incorporated 5,6-spirocycle rigid structural motifs into the linker, yielding multiple benefits: Rigid structures reduce the number of flexible rotatable bonds, lowering conformational entropy penalties and improving oral absorption - Ether rings provide additional polar surface area, enhancing aqueous solubility - Achieved an optimal balance between rigidity and polarity, overcoming the traditional challenge of poor water solubility in large-molecule drugs.
Ultimately, spirocyclic linker variants achieved a >100-fold improvement in solubility (FeSSIF, Fasted State Simulated Intestinal Fluid) over flexible linkers (reaching 156 μM), with corresponding improvements in oral bioavailability.
2.3.2
Round 2: MCS Privileged Fragment Guidance and Rigidification
Round 2: MCS Privileged Fragment Guidance and Rigidification
During the linker optimization phase, the following additional Chemistry42 built-in tools were utilized:
- Alchemistry: Performed relative binding free energy (RBFE) calculations on shortlisted candidates during the warhead optimization stage, where the compounds are conventional small molecules, and RBFE predictions are well-established and reliable. For PROTAC-sized molecules, however, the accuracy of RBFE calculations remains an area of active investigation; the Alchemistry results at the PROTAC stage were therefore used as directional guidance to complement, rather than replace, experimental affinity data. Further experimental work is needed to fully validate the accuracy of RBFE for macromolecular PROTACs.
- Retrosynthesis: Assessed the synthetic feasibility of generated molecules, ensuring that AI-designed structures possess reasonable synthetic routes and avoiding “computationally perfect but unsynthesizable” candidates.
3. Workflow Output: Final Molecule ISM2286
Complete in vitro and in vivo experimental data, proteomic validation, and pharmacokinetic analyses are available in the published manuscript. Below is a brief summary of key validation results for the Chemistry42 workflow output molecule ISM2286 (D16-M1P2).
Figure 13.
Overall structural schematic of the final candidate molecule ISM2286 (D16-M1P2), showing the three-module architecture of warhead (ISM4963), spirocyclic linker, and CRBN binder, along with key protein interactions.
4. Conclusion
Using the development of ISM2286, a PKMYT1 PROTAC degrader published in Nature Communications, as a case study, this white paper provides a detailed account of how the Chemistry42 platform was applied at each critical stage of PROTAC development, including operational procedures and design strategies.
The core value of Chemistry42 lies not only in the technical capability of individual modules but in integrating meticulous input configuration, iterative configuration optimization, AI generation, selectivity prediction, molecular dynamics simulation, and ADMET optimization into an end-to-end, controllable, and reproducible industrialized design loop. In the PROTAC field, this workflow paradigm organically combines drug discovery experts' structural insight and manual refinement, significantly reducing R&D uncertainty and accelerating the discovery of high-quality candidate molecules.
The core value of Chemistry42 lies not only in the technical capability of individual modules but in integrating meticulous input configuration, iterative configuration optimization, AI generation, selectivity prediction, molecular dynamics simulation, and ADMET optimization into an end-to-end, controllable, and reproducible industrialized design loop. In the PROTAC field, this workflow paradigm organically combines drug discovery experts' structural insight and manual refinement, significantly reducing R&D uncertainty and accelerating the discovery of high-quality candidate molecules.
References
- Wang, Y., Wang, X., Liu, T. et al. Discovery of a bifunctional PKMYT1-targeting PROTAC empowered by AI-generation. Nature Communications 16, 10759 (2025).
- Gallo, D., Young, J. T. F., Fourtounis, J. et al. CCNE1 amplification is synthetic lethal with PKMYT1 kinase inhibition. Nature604, 749–756 (2022).
- Szychowski, J., Papp, R., Dietrich, E. et al. Discovery of an orally bioavailable and selective PKMYT1 inhibitor, RP-6306. Journal of Medicinal Chemistry 65, 10251–10284 (2022).
Appendix
This white paper was prepared by Insilico Medicine. For further information about Chemistry42’s applications in targeted protein degradation, please contact us. info@insilico.com
Stay tuned, follow us on social media!