Benchmarking 3D Structure-Based Molecule Generators

Area of Science:

• * Computational chemistry and cheminformatics.
• * Artificial intelligence in drug discovery.
• * Molecular modeling and simulation.

Background:

• * Evaluating generative models for drug design is crucial.
• * Existing methods for generating 3D molecules have limitations.
• * Protein-ligand interactions and conformations are key drug design targets.

Purpose of the Study:

• * To benchmark 3D combinatorial and deep learning generators.
• * To assess their ability to recreate protein-ligand interactions and conformations.
• * To identify strengths and weaknesses of different generative approaches.

Main Methods:

• * Developed a novel benchmark using the BindingMOAD dataset.
• * Evaluated sequential graph neural networks (Pocket2Mol, PocketFlow), diffusion models (DiffSBDD, MolSnapper), and genetic algorithms (AutoGrow4, LigBuilderV3).
• * Assessed structural validity, 3D ligand conformations, and interaction recreation.

Main Results:

• * Deep learning models failed to generate structurally valid molecules and conformations.
• * Combinatorial methods were slow and produced molecules failing 2D filters.
• * Identified specific limitations of both deep learning and combinatorial approaches.

Conclusions:

• * Deep learning generators need improved focus on structural validity and interaction accuracy.
• * Combinatorial generators require optimization for speed and filter compliance.
• * The benchmark provides a framework for advancing structure-based drug design generators.