Autoregressive fragment-based diffusion for pocket-aware ligand design

TL;DR

This work addresses pocket-aware 3D ligand design by introducing AutoFragDiff, a fragment-based autoregressive diffusion model conditioned on protein pockets and molecular scaffolds. It leverages Geometric Vector Perceptrons and an order-agnostic diffusion objective to generate fragments with improved local geometry, while enabling scaffold extension and anchor-point prediction via a dedicated AnchorGNN. Evaluations on the CrossDock dataset show improved local structural fidelity (lower JSD for angles/dihedrals) and higher predicted binding affinity, with scaffold extensions frequently enhancing potency. The approach provides a practical, open-source framework for pocket-aware lead optimization, with future directions including affinity-guided guidance and architectural refinements to further boost performance.

Abstract

In this work, we introduce AutoFragDiff, a fragment-based autoregressive diffusion model for generating 3D molecular structures conditioned on target protein structures. We employ geometric vector perceptrons to predict atom types and spatial coordinates of new molecular fragments conditioned on molecular scaffolds and protein pockets. Our approach improves the local geometry of the resulting 3D molecules while maintaining high predicted binding affinity to protein targets. The model can also perform scaffold extension from user-provided starting molecular scaffold.

Figures (9)

• Figure 1: Noising and Sampling for a single fragment inside a protein pocket. Yellow spheres show the anchor point.
• Figure 2: (Left) The fragment connectivity for a molecule. Highlighted atoms are the anchor points on each fragment. (Right) Sampled generation order for the molecule on the left from its fragments.
• Figure 3: Scaffold (red) extension examples on a Cytochrome-c peroxidase (pdb: 1a2g).
• Figure 4: Steric clashes of different models.
• Figure 5: Strain energies of different models