Fragment and Geometry Aware Tokenization of Molecules for Structure-Based Drug Design Using Language Models.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• Structure-based drug design (SBDD) is vital for targeted therapeutics but faces challenges in modeling complex protein-ligand interactions and exploring vast chemical spaces.
• Language models (LMs), successful in natural language processing, have been minimally explored for SBDD applications.

Purpose of the Study:

• To introduce Frag2Seq, a novel method applying LMs to SBDD for efficient and effective molecule generation.
• To enhance target-aware molecule generation by incorporating protein pocket information into LMs.

Main Methods:

• Frag2Seq transforms 3D molecules into fragment-informed sequences using SE(3)-equivariant frames, preserving geometric data.
• Protein pocket embeddings from an inverse folding model are integrated into LMs via cross-attention to capture protein-ligand interactions.
• The method utilizes fragment-based generation and protein context encoding for molecule design.

Main Results:

• Frag2Seq achieved superior performance in binding affinity (vina score) and drug-likeness (QED, Lipinski rules).
• The model demonstrates efficacy in generating ligands with enhanced binding affinity to target proteins.
• Frag2Seq offers significantly higher sampling efficiency compared to existing autoregressive and diffusion models, with up to a 300x speedup.

Conclusions:

• Frag2Seq effectively leverages LMs for SBDD by combining fragment-based generation with protein context.
• The method shows promise for accelerating the discovery of specific and effective drug candidates.
• Frag2Seq represents a significant advancement in computational approaches for drug discovery, offering both accuracy and speed.