Insilico Medicine published a white paper in June 2026 describing how its Chemistry42 generative chemistry platform was used at each stage of designing ISM2286, a first-in-class bifunctional PROTAC degrader of the kinase PKMYT1. The experimental validation appeared in Nature Communications (Wang et al., volume 16, article 10759, 2025).
PKMYT1 regulates the cell cycle and is synthetic lethal with CCNE1-amplified tumours. Existing inhibitors such as RP-6306 validated the target but rely on sustained occupancy of the ATP site, with the usual problems of selectivity, resistance and therapeutic window. Insilico chose a PROTAC approach, which recruits an E3 ligase to degrade PKMYT1, and rather than converting RP-6306 into a degrader it set out to design a new warhead with better kinase selectivity, a linker-accessible binding orientation and a PROTAC-compatible physicochemical profile.
For the warhead, the team used RP-6306 as the reference ligand, set the two hinge-region hydrogen bonds as mandatory pharmacophore points, fixed the phenol region as a privileged fragment, removed crystallographic water facing the solvent, applied a 5 angstrom pocket dilation and fused in solvent-exposed features of dasatinib. More than 20 generative models proposed scaffolds, which were filtered by pocket-ligand interaction scoring with ligand efficiency metrics as references, over 900 medicinal chemistry filters, 140 ADMET models and Golden Cubes, a module that predicts the probability of inhibiting each kinase; only molecules with five or fewer predicted off-targets advanced. The paper says Golden Cubes compresses "what traditionally requires weeks of wetlab experimental screening for kinome selectivity into a matter of hours."
Medicinal chemists then refined the AI-generated pyrrolo[2,3-b]pyridine scaffold by adding a cyano group, giving Compound 3. Its pharmacokinetics were poor (clearance of 63.6 mL/min/kg and 6% oral bioavailability in rats), and the ADMET tools had flagged a pyrimidine methyl group as an oxidative hotspot, so it was replaced with an amine, giving the final warhead ISM4963. Its amine tail sits close to the solvent, allowing a short linker of about 8 angstroms. In experiments the number of off-target kinases fell from 46 for RP-6306 to 11 for the new warhead, a reduction of more than 76%.
CRBN was chosen as the E3 ligase and its binder was selected manually by medicinal chemists. External docking tools and molecular dynamics were used to pick a ternary complex model, and an E2 enzyme from experimental structures or AlphaFold3 predictions was added to check whether PKMYT1 residue Lys345 faces E2 in a geometry suited to ubiquitination; the paper notes this remains a computational hypothesis awaiting experimental validation.
For the linker, Anchor Points locked both binders while the generative models explored the space between them under the 8 angstrom constraint. An unconstrained first round starting from a basic PEG linker surfaced piperazine-containing linkers that form a hydrogen bond with PKMYT1 residue Leu116; the resulting hit, D1, had a DC50 of 1 nM but poor solubility and high plasma protein binding. A second round used maximum common substructures from public PROTAC linker libraries as privileged fragments and enumerated spirocyclic motifs, and the resulting 5,6-spirocyclic linkers improved solubility in simulated intestinal fluid more than 100-fold, to 156 micromolar. Relative binding free energy calculations and retrosynthesis tools were also used.
The final molecule, ISM2286, was positive against only 4 of 403 kinases, including PKMYT1, and whole-proteome analysis showed PKMYT1 as the only significantly downregulated protein. It combined degradation with inhibition, keeping CDK1 phosphorylation suppressed for more than 24 hours after washout, showed oral bioavailability of 91.3% in mice, 166.5% in rats, 67.9% in dogs and 20.5% in monkeys, and achieved 66.4% tumour growth inhibition with oral dosing in a CCNE1-amplified xenograft model.