AlphaFold Meets\nDe Novo Drug Design: Leveraging Structural\nProtein Information in Multitarget Molecular Generative Models

Recent advancements in deep learning and generative models\nhave\nsignificantly expanded the applications of virtual screening for drug-like\ncompounds. Here, we introduce a multitarget transformer model, <i>PCMol</i>, that leverages the latent protein embeddings derived\nfrom <i>AlphaFold2</i> as a means of conditioning a de novo\ngenerative model on different targets. Incorporating rich protein\nrepresentations allows the model to capture their structural relationships,\nenabling the chemical space interpolation of active compounds and\ntarget-side generalization to new proteins based on embedding similarities.\nIn this work, we benchmark against other existing target-conditioned\ntransformer models to illustrate the validity of using AlphaFold protein\nrepresentations over raw amino acid sequences. We show that low-dimensional\nprojections of these protein embeddings cluster appropriately based\non target families and that model performance declines when these\nrepresentations are intentionally corrupted. We also show that the <i>PCMol</i> model generates diverse, potentially active molecules\nfor a wide array of proteins, including those with sparse ligand bioactivity\ndata. The generated compounds display higher similarity known active\nligands of held-out targets and have comparable molecular docking\nscores while maintaining novelty. Additionally, we demonstrate the\nimportant role of data augmentation in bolstering the performance\nof generative models in low-data regimes. Software package and AlphaFold\nprotein embeddings are freely available at https://github.com/CDDLeiden/PCMol.