AlphaFold Meets De Novo Drug Design: Leveraging Structural Protein Information in Multitarget Molecular Generative...

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Deep learning and generative models have advanced virtual screening for drug-like compounds.
• Conditioning generative models on protein targets is crucial for de novo drug design.

Purpose of the Study:

• Introduce PCMol, a multitarget transformer model for de novo drug generation.
• Leverage AlphaFold2 protein embeddings to condition generative models on specific protein targets.
• Evaluate the effectiveness of protein representations in target-conditioned drug discovery.

Main Methods:

• Developed a multitarget transformer model (PCMol) utilizing latent protein embeddings from AlphaFold2.
• Conditioned the de novo generative model on diverse protein targets using their embeddings.
• Benchmarked PCMol against existing transformer models using raw amino acid sequences.
• Analyzed clustering of protein embeddings and model performance with corrupted representations.
• Demonstrated the impact of data augmentation on generative model performance in low-data scenarios.

Main Results:

• PCMol effectively captures protein structural relationships, enabling chemical space interpolation and target generalization.
• AlphaFold protein representations outperform raw amino acid sequences for target-conditioned generation.
• Protein embeddings show appropriate clustering by target families, and performance degrades with corrupted embeddings.
• PCMol generates diverse, potentially active molecules for various proteins, including those with sparse data.
• Generated compounds exhibit higher similarity to known ligands, comparable docking scores, and maintained novelty.
• Data augmentation significantly improves generative model performance in low-data regimes.

Conclusions:

• PCMol offers a novel approach to de novo drug design by integrating structural protein information.
• The use of AlphaFold2 embeddings enhances the accuracy and generalization capabilities of target-conditioned generative models.
• The findings highlight the importance of rich protein representations and data augmentation for effective virtual screening and drug discovery.