De novo generation of multi-target compounds using deep generative chemistry

Classical drug discovery follows a "one disease–one target–one drug" model. While successful, this model is limited for diseases with multiple molecular causes, such as cancer and psychiatric disorders. These polygenic diseases involve complex biological networks and multiple potential intervention points. This necessitates a shift towards multi-target treatment strategies, achievable through combination therapies or polypharmacology drugs, which simultaneously modulate two or more targets. While still nascent, polypharmacology demonstrates efficacy, particularly in treating previously recalcitrant cancers like KRAS mutant non-small cell lung cancers. Polypharmacology offers advantages over combination therapies, including improved pharmacokinetic and safety profiles, reduced resistance development, and simplified treatment regimens. A major hurdle in polypharmacology is designing a single agent potently inhibiting multiple proteins. Past successes, such as targeting RET and VEGFR2 in thyroid cancer, have been serendipitous rather than systematic, requiring substantial time and resources. Recent machine learning advancements show promise in predicting compound-target interactions, generating single-target inhibitors, and identifying existing drugs with polypharmacology potential. This paper aims to address the systematic generation of polypharmacology compounds through a novel approach.

The authors review existing literature on polypharmacology, highlighting its advantages and challenges. They cite studies demonstrating the successful application of polypharmacology in treating complex diseases, particularly cancer. The limitations of traditional drug discovery methods in addressing polygenic diseases are discussed, emphasizing the need for new approaches. The authors also acknowledge recent progress in machine learning for drug discovery, including predicting compound-target interactions and generating de novo single-target inhibitors. However, they point out the lack of systematic methods for designing polypharmacology compounds, motivating the development of their novel approach.

The researchers developed POLYGON, a deep machine learning model based on generative AI and reinforcement learning. POLYGON employs a variational autoencoder (VAE) to create a chemical embedding, a low-dimensional representation of chemical space. The VAE, trained on over one million small molecules from the ChEMBL database, encodes chemical formulas into embeddings and decodes embeddings back into valid formulas. The model’s effectiveness was validated by its ability to encode and recover chemical formulas of held-out molecules and to generate valid SMILES strings from most coordinates in the embedding. Importantly, compounds with affinity for the same target were significantly closer in the chemical embedding than those with affinity for different targets. The embedding framework underpins a reinforcement learning system, iteratively sampling compounds from the embedding and rewarding those with predicted inhibitory ability against two targets, alongside desirable drug-likeness and synthesizability properties. The reinforcement learning system refocuses the chemical embedding in successive epochs, progressively improving compound quality. The accuracy of the compound-target scoring module was benchmarked against a held-out set of compounds with characterized IC50s against dual targets, achieving 81.9% accuracy at an activity threshold of IC50 < 1 µM. Further, the scoring module was compared favorably to existing methods in a recent IDG-DREAM competition. Subsequently, POLYGON was used to generate de novo compounds targeting ten pairs of synthetically lethal cancer proteins, with docking analysis indicating favorable binding energies for the top-scoring compounds. Finally, 32 compounds targeting MEK1 and mTOR were synthesized and validated through cell-free assays and lung tumor cell experiments.

POLYGON accurately identifies polypharmacology interactions with 82.5% accuracy. De novo compounds generated by POLYGON targeting ten pairs of synthetically lethal cancer proteins showed favorable binding energies in docking analysis, indicating strong binding to their intended targets. The 32 synthesized MEK1/mTOR inhibitors exhibited significant reductions (over 50%) in both kinase activities and cell viability at doses of 1–10 µM. Four compounds demonstrated greater than 50% reduction in both mTOR and MEK1 phosphorylation activity at 1 µM. The observed reductions in mTOR and MEK1 activity strongly correlated with overall growth inhibition. In-vitro kinase assays confirmed the inhibitory capacity of the lead compound IDK12008, with IC50 values consistent with those observed in human cells. Off-target analysis indicated minimal inhibition of unrelated kinases, suggesting good specificity, although a degree of promiscuity similar to FDA-approved kinase inhibitors was observed. The AIWIC model, a top performer in the DREAM challenge, also predicted low dissociation constants for a majority of the generated compounds for at least one of the targets, and for 20% of the compounds, it predicted low dissociation constants for both targets.

This study demonstrates the successful application of generative reinforcement learning for the design of polypharmacology drugs. The POLYGON model systematically generates, synthesizes, and validates compounds with activity against two targets. This addresses a critical gap in polypharmacology drug design, moving beyond serendipitous discovery. The high accuracy of POLYGON in identifying and generating dual-target inhibitors validates its potential for accelerating drug development. The successful experimental validation of synthesized compounds supports the predictive power of the model and demonstrates the feasibility of this approach. The observed correlation between inhibition of mTOR and MEK1 and cell growth reduction highlights the importance of targeting multiple pathways in cancer therapy. However, further optimization using structure-activity relationships (SAR) and inclusion of ADMET properties is warranted.

POLYGON provides a novel pipeline for designing, synthesizing, and validating polypharmacology compounds. The high success rate in generating and validating MEK1/mTOR inhibitors showcases the potential of generative modeling for drug discovery. Future research should focus on incorporating ADMET properties, improving the model's selectivity, and exploring additional synthetic lethal combinations to further expand the utility of this approach. Integrating SAR data from synthesized compounds into future training will likely enhance potency and selectivity.

The current study focused primarily on kinase targets and did not extensively investigate ADMET properties. While the off-target analysis showed minimal effects on some representative kinases, complete off-target profiling might reveal additional interactions. The study's scope is limited to a small set of synthesized compounds, warranting further investigation of a broader range of molecules. The number of synthesized MEK1/mTOR compounds (32 out of 100) limits the scope of the experimental validation. The generalizability of the model to other target pairs beyond those tested needs further exploration.