Munson et al. recently reported POLYGON, a model for de novo generation of compounds that inhibit multiple proteins. POLYGON employs generative AI and reinforcement learning to design novel structures by sampling from an embedded chemical space. The rewarded, desired qualities are predicted inhibitory potential against each of two protein targets as well as drug-likeness and synthetizability.
Traditionally, drug discovery has focused on inhibitors targeting single proteins. While effective, this approach may not be ideal for all diseases. Some conditions might benefit more from targeting multiple proteins that contribute partially to the illness. This has led to a growing interest in developing multi-targeted therapies.
One approach is to use combination therapy (combining multiple drugs), and another is the so-called polypharmacology, which involves the use of a single drug that modulates multiple targets simultaneously. Although polypharmacology is an interesting strategy with many potential advantages over combination therapy, the challenge of designing such multi-target drugs has been a major limiting factor.
In a benchmark comprising 109,811 compounds (1850 targets) from the BindingDB data set, POLYGON was able to classify compounds with activity on two targets at an accuracy of ~82%. Further, POLYGON was applied to generate compounds targeting ten pairs of known co-dependent (i.e. synthetic lethal) protein targets including Ser/Thr kinases, Tyr kinases, DNA binding factors, and histone modifiers. The top predicted structures showed low binding energy to their corresponding pairs of targets according to molecular docking analysis. In addition, the top structures showed comparable 3D poses to known single-target inhibitors.
The authors synthesized and experimentally tested 32 POLYGON-generated de novo compounds designed for dual inhibition of MEK1 and mTOR. Most of the compounds had IC50s in the 1–10 μM range and lowered the in vitro and cellular (at 1-10 μM) activity of each target by more than 50%. Further, four of the well-validated POLYGON compounds were evaluated for specificity by assessing off-target inhibition of three representative unrelated kinases. All the compounds showed no more than 20% inhibition except for one compound with 38% inhibition towards one off-target kinase.
Open questions
❓ Can this approach be used for the discovery of PROTACs and molecular glues, e.g. to recruit an E3 ligase to a protein of interest?
❓How well will POLYGON translate to other drug targets and disease areas that have different protein interactions and functionalities?
References and Resources
Paper: De novo generation of multi-target compounds using deep generative chemistry
Code: POLYpharmacology Generative Optimization Network (POLYGON) a VAE for de novo polypharmacology.
