In a new study published in Nature Communications, our team of researchers from Insilico Medicine demonstrate how generative AI can solve long-standing pharmacological challenges. By designing a novel, highly selective PKMYT1 inhibitor and integrating it into a bifunctional PROTAC (Proteolysis Targeting Chimera), we developed D16-M1P2, a molecule that not only inhibits but actively degrades the target, offering a more potent, more targeted, and potentially safer therapeutic strategy.
Synthetic Lethality to Target Cancers
Cell cycle dysregulation is a hallmark of cancer, driving uncontrolled proliferation and genomic instability. While this chaos fuels tumor growth, it also creates specific vulnerabilities. Synthetic lethality is a scenario where a targeted therapeutic works together with underlying disrupted biological systems to specifically target disease-causing cells. For cancers driven by mutations in genes responsible for controlling the cell cycle, targeting other genes that also regulate the cell cycle– which the cell now relies more heavily on– could send the cancer into a self-destructive cycle of increasing genomic instability.
The kinase PKMYT1 has emerged as one such critical guardian of the cell cycle. PKMYT1 limits the activity of CDK1, an essential cell cycle factor that triggers cells to progress into mitosis. A number of common cancer-causing mutations, such as CCNE1 amplification or deleterious mutations in FBXW7 and PPP2R1A, converge on CDK1 to modulate the cell cycle. Inhibiting PKMYT1, which then unleashes CDK1 signaling, forces these cancer cells into premature mitosis before their genomes are properly replicated, leading to catastrophic instability and cell death.
However, targeting cell cycle regulators and, more generally, kinases has historically been fraught with difficulty. Selective inhibitors are hard to design, often hitting off-target proteins and other kinases, which leads to dose-limiting toxicities and limited clinical benefit. We needed a drastically different approach to develop a safe and effective kinase inhibitor, one which new AI tools can uniquely facilitate.
The Challenge: Designing a Better Binder
The current leading clinical-stage PKMYT1 inhibitors bind deep within the active site pocket of the protein. Although potent inhibitors of PKMYT1 kinase activity, they suffer from two major issues. First, insufficient selectivity from off-target inhibition of important kinases like BRAF and SRC contribute to toxicities. Second, PKMYT1 can activate Wnt/β-catenin signaling through mechanisms outside its kinase activity, thereby contributing to a cancer cell’s oncogenicity, so fully inhibiting PKMYT1 requires a new strategy beyond inhibition of its enzymatic site.
To overcome these challenges, we decided to use Insilico’s advanced AI-based drug design platforms to develop a compound highly specific to PKMYT1 that could also be converted into a PROTAC, a type of engineered molecule that helps degrade target proteins by bringing together the target with the cell’s endogenous protein-recycling machinery (a ubiquitin ligase) to mark the protein for destruction. This bold goal was only more ambitious because a PKMYT1 degrader had never been developed before.
The existing PKMYT1 inhibitors are poor candidates for conversion into a PROTAC, because a target-binding domain buried too deep in the pocket makes attaching a linker segment that joins the PROTAC’s target-binder and ligase-binder domains chemically difficult and often destroys the molecule’s drug-like properties. We determined that an entirely new PKMYT1 inhibitor was needed, which could not only bind with high specificity to the binding pocket but also extend outward into the solvent-exposed area to offer a more accessible “tethering point” to attach the rest of the PROTAC structure.
AI-Driven Generative Chemistry: Solving the Selectivity Puzzle
Using the Chemistry42 generative AI platform, we employed a fragment-based design strategy. We anchored the pocket-binding pharmacophore of a known PKMYT1 inhibitor along with the solvent-extending features of the broad-spectrum inhibitor dasatinib.
The AI platform then generated over 2,000 novel structures. Through iterative filtering, which included clustering, docking, and visual inspection, fragment-based recombination, and targeted and systematic modification of key atoms and functional groups, we pieced together a novel molecular structure of an entirely new chemotype. Crucially, we utilized the Golden Cubes kinase toolkit within Chemistry42 to predict and minimize off-target effects.
This process yielded Compound 4, a novel inhibitor featuring a pyrrolopyridine heterocycle, which enables entirely new binding interaction distinct from previously reported PKMYT1 inhibitors, and nanomolar-level inhibitory capacity against PKMYT1 and downstream CDK1 signaling. Most notably, Compound 4 represents what is likely the first AI-generated small molecule with such exceptional kinase selectivity, validating our AI-based toolset to engineer out drug target promiscuity.
Because Compound 4 extends into the solvent front, it provided the perfect "handle" for constructing a PROTAC.
From Inhibitor to Degrader: Engineering PROTAC D16-M1P2
Engineering a PROTAC requires integrating three distinct domains into one molecule: (1) a target-binding domain, (2) a ubiquitin ligase-binding domain, and (3) a linker to join them.
The target-binding domain was achieved with Compound 4, and we had settled on linking Compound 4 to a known ligand for cereblon (CRBN), an E3 ubiquitin ligase widely expressed in tumors that acts as the target for a number of promising clinical-stage PROTACs. To design the linker segment, we 3D modeled PKMYT1 in proximity to CRBN to determine that an 8 Å distance was required to bring the two into the proximity necessary for ubiquitin transfer. Chemistry42 was again deployed to generate linker structures bridging Compound 4 and the CRBN-binder.
The initial design, Compound D1, showed promise but suffered from poor solubility and other characteristics commonly challenging to PROTAC development. Through systematic optimization of the linker and CRBN-binder domains, we discovered a key modification: the inclusion of a 5,6-spirocycle in the linker. This rigid spirocycle reduced the number of rotatable bonds, locking the molecule into a conformation favorable for forming the target ternary complex. This single modification led to a 3-fold improvement in cellular potency, significantly higher oral exposure, and ~50% greater bioavailability in animal models.
Dual Mechanism: The Power of Degradation
The resulting lead candidate, D16-M1P2, now represents the first PKMYT1 degrader, functioning via the dual mechanisms of inhibiting enzymatic activity of PKMYT1 and recruiting cellular machinery to eliminate the protein. Along with an impressive PKMYT1 degradation DC50 of 0.7 nM, cellular CDK1 phosphorylation was inhibited at IC50 of 9.0 nM, representing a highly potent inhibitor of PKMYT1.
Validation studies revealed that degradation is the dominant driver of its potency. In washout experiments, where the drug is removed after treatment, PKMYT1 activity bounced back within two hours of treatment with the established PKMYT1 inhibitor. In contrast, D16-M1P2 maintained PKMYT1 suppression for over 24 hours. This durable, "event-driven" pharmacology means the drug continues to work even after it stops binding the target, as the protein has already been marked for destruction.
Furthermore, D16-M1P2 demonstrated a massive improvement in specificity, even beyond that of Compound 4. In kinase profiling, the PROTAC inhibited only 4 kinases to a notable degree (including PKMYT1), but exhibiting approximately 50-fold selectivity for PKMYT1 against the next-strongest off-target hits like BRAF and SRC. By comparison, the established PKMYT1 inhibitor showed less than 10-fold selectivity and over ten times as many off-target hits.
Validation in Oncology Models
As expected from the mechanism of PKMYT1 inhibition and its role in controlling cell cycle progression, D16-M1P2 exhibited substantial synthetic lethality with cell cycle-disrupting oncogenic mutations against tumor models in vitro and in vivo. The PROTAC displayed a 12-fold greater selectivity for cancer cells with CCNE1 amplification or FBXW7/PPP2R1A mutations compared to wild-type cells.
The superior selectivity of D16-M1P2 also translated into a wider therapeutic window in vivo. In mouse xenograft models, D16-M1P2 treatment induced none of the significant body weight loss associated with treatment with the established PKMYT1 inhibitor, potentially due to its much reduced off-target inhibition of other important kinases.
Optimization of drug-like properties using our AI platforms at every step of the discovery process, from Compound 4 structure to linker design, led to very promising pharmacokinetic properties indicative of effectiveness as a drug. Across preclinical species, D16-M1P2 showed excellent oral absorption, bioavailability, and exposure, and low clearance, all supporting its potential for oral drug administration.
Final Takeaways
This study highlights several milestones in AI-driven drug discovery we're proud to showcase. The work demonstrates perhaps the earliest success story in using AI to generate a small molecule (Compound 4) designed specifically to overcome the kinase selectivity issues that plague traditional discovery, thanks to the dedicated kinase-selectivity Golden Cubes analytical toolkit. The novel scaffold integrated into Compound 4 binds both deeply within the pocket and at the solvent front, allowing for the creation of a dual inhibitor-degrader PROTAC. It also demonstrates the flexibility of the Chemistry42 platform, extending its capabilities beyond small molecules to complex, multi-domain modalities like PROTACs.
While PROTACs face inherent challenges, many of which are different than those facing small-molecule drugs, D16-M1P2 represents a significant leap forward. It serves as both a precise chemical probe for understanding PKMYT1 biology and a promising lead candidate for distinct, safer, and more durable cancer therapy.