Unraveling the Potential of Diffusion Models in Small Molecule Generation

A review of diffusion models for molecular generation and drug discovery.

Scope

This review studies how diffusion models are being used for small-molecule generation and drug discovery, from basic mathematical formulations to protein-pocket-aware design.

It explains forward and reverse diffusion, DDPMs, score-based models, and the role of equivariance in preserving the geometry of 3D molecules.

A taxonomy of molecular diffusion models

The paper organizes methods along several dimensions:

  • target-free versus target-aware generation;
  • molecular conformation generation versus de novo molecule generation;
  • 1D SMILES, 2D graphs, and 3D point-cloud or voxel representations;
  • DDPM versus score-based formulations; and
  • SE(3)-equivariant, permutation-equivariant, or non-equivariant architectures.

This taxonomy connects the representation and conditioning choices to practical goals such as chemical-space exploration, conformation generation, docking, and ligand design.

Benchmark and findings

The study benchmarks 18 representative models across QM9, GEOM-Drugs, and CrossDocked2020. It considers validity, uniqueness, novelty, atom and molecule stability, Validity3D, QED, synthetic accessibility, Vina score, and strain energy.

For target-free generation, MiDi shows a strong combination of stability and validity. For target-aware generation, KGDiff performs best in the post-relaxation comparison, although results vary substantially across models and targets.

The review also finds that force-field relaxation can increase average 3D validity from 1.42% to 37.0%, while sometimes changing binding-related scores. This highlights the need for evaluation protocols that measure both chemical validity and physical realism.

Open challenges

The paper identifies data sparsity, limited interpretability, unreliable evaluation metrics, high computational cost, and missing physics-based constraints as major obstacles.

Future directions include physics-informed diffusion, better target-aware and multimodal modeling, more realistic benchmarks, and methods that make the generation process easier to inspect.

Paper · arXiv