A 3D pocket-aware and evolutionary conserved interaction guided diffusion model for molecular optimization

Anjie Qiao, Hao Zhang, Qianmu Yuan, Qirui Deng, Jingtian Su, Weifeng Huang, Huihao Zhou, Guo-Bo Li, Zhen Wang, Jinping LeiarXiv 2025 · DOI · arXiv

Introduction

In structure-based drug design (SBDD), the primary objective is to generate molecules that bind effectively to a specific protein pocket. While modern diffusion models have demonstrated proficiency in generating molecules that conform to 3D pocket geometries, they often fail to distinguish between residues of varying biological importance. Evolutionarily conserved residues are typically vital for a protein's function, yet standard AI models often treat all residues as equal. This can result in molecules that show high binding scores in simulations but lack biological activity in experimental settings.

This paper introduces DiffDecip, a 3D target-aware diffusion model designed to bridge this gap. By integrating evolutionary conservation data and pretrained interaction priors, DiffDecip generates drug candidates that prioritize interactions with the most biologically critical residues of a protein pocket.

Methods

DiffDecip is built upon DiffDec, an E(3)-equivariant 3D conditional diffusion model. The authors introduced two primary innovations to guide the molecular generation process:

  1. Conservation-Aware Conditioning (Evo-guidance): The model utilizes Multiple Sequence Alignment (MSA) to calculate a "conservation score" for every residue in the target protein pocket. This score serves as an additional feature, highlighting functional "hotspots" where the molecule should ideally form interactions.
  2. Interaction-Prior Guidance: The system integrates a pretrained network called IPNet. During the diffusion process, where the model denoises a random cloud of atoms into a structured molecule, IPNet provides a mathematical "mean shift." This nudge steers atoms toward positions that facilitate favorable non-covalent interactions, such as hydrogen bonds or hydrophobic contacts.

The model focuses on scaffold decoration, a task involving the addition of new chemical groups (R-groups) to a known molecular core to optimize its pharmacological properties.

The architecture of DiffDecip integrates Evo-guidance from MSA and Interaction Guidance from IPNet to steer the diffusion process toward biologically relevant molecular configurations.

Experimental Setup

The researchers utilized the CrossDocked dataset, comprising 22.5 million docked protein-ligand complexes. The data was filtered for high-quality poses (RMSD < 1Å) and protein diversity (sequence identity < 30%). Using the LibINVENT tool, ligands were segmented into scaffolds and R-groups based on 37 reaction rules, resulting in a training set of 76,370 tuples.

Results and Discussion

DiffDecip was evaluated against the baseline DiffDec model, showing significant improvements in both binding affinity and biological relevance:

  • Binding Affinity: DiffDecip achieved a superior average Vina score of -8.19, compared to -8.11 for the baseline. It outperformed the baseline in 65.3% of test cases.
  • Targeting Conserved Residues: This was the model's most notable strength. DiffDecip formed more interactions with highly conserved residues in 84% of the protein targets.
  • Chemical Validity and Uniqueness: The model maintained a high validity rate of 90.1%. Interestingly, the inclusion of biological priors increased the uniqueness of generated molecules from 53.3% to 67.3%, suggesting that these constraints help the model explore more diverse and valid chemical spaces.
Performance comparison between DiffDecip and the baseline DiffDec. (b) Average Vina scores across cases. (c, d) Average interaction numbers with conserved residues, showing DiffDecip's superior ability to target functional hotspots.

Conclusion

DiffDecip demonstrates that prioritizing "functional fit" over simple geometric complementarity leads to more biologically relevant drug candidates. By incorporating evolutionary conservation and interaction priors, the model effectively targets critical residues within protein pockets. Future research will focus on extending this guidance to multi-site scaffold modification and de novo drug design.

Summary updated

BibTeX
@misc{qiao2025d,
  author = {Anjie Qiao and Hao Zhang and Qianmu Yuan and Qirui Deng and Jingtian Su and Weifeng Huang and Huihao Zhou and Guo-Bo Li and Zhen Wang and Jinping Lei},
  title = {{A 3D pocket-aware and evolutionary conserved interaction guided diffusion model for molecular optimization}},
  year = {2025},
  eprint = {2505.05874},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.05874},
}
Figure