AUTODIFF: Autoregressive Diffusion Modeling for Structure-based Drug Design

TL;DR

Autodiff tackles structure-based drug design by formulating conditional 3D molecule generation $p(\mathcal{G}|\mathcal{P})$ conditioned on protein pockets and introducing conformal motifs to preserve local 3D topology. It combines a fragment-wise, diffusion-based generation with an $SE(3)$-equivariant encoder to model pocket–ligand interactions, and employs a focal-connection-site framework (with RCCS) to attach motifs and diffuse torsions. The approach yields realistic molecular structures with valid bond angles and competitive binding affinities, validated on CrossDocked2020 under molecular weight constraints and supported by new structure- and affinity-related metrics. Overall, AutoDiff advances SBDD by integrating geometry-aware motif design, diffusion-based torsion modeling, and fair evaluation, enabling practical generation of high-quality drug candidates.

Abstract

Structure-based drug design (SBDD), which aims to generate molecules that can bind tightly to the target protein, is an essential problem in drug discovery, and previous approaches have achieved initial success. However, most existing methods still suffer from invalid local structure or unrealistic conformation issues, which are mainly due to the poor leaning of bond angles or torsional angles. To alleviate these problems, we propose AUTODIFF, a diffusion-based fragment-wise autoregressive generation model. Specifically, we design a novel molecule assembly strategy named conformal motif that preserves the conformation of local structures of molecules first, then we encode the interaction of the protein-ligand complex with an SE(3)-equivariant convolutional network and generate molecules motif-by-motif with diffusion modeling. In addition, we also improve the evaluation framework of SBDD by constraining the molecular weights of the generated molecules in the same range, together with some new metrics, which make the evaluation more fair and practical. Extensive experiments on CrossDocked2020 demonstrate that our approach outperforms the existing models in generating realistic molecules with valid structures and conformations while maintaining high binding affinity.

Figures (6)

• Figure 1: Illustration of the advantage of conformal motif (bottom) versus other methods Zhang2022_FLAGZhang2023_DrugGPS (top).
• Figure 2: Overview of the generation process of AutoDiff. RCCS: Reduced Candidate Connection Site. FCS: Focal Connection Site. Details are shown in \ref{['tiffmlp']}.
• Figure 3: Comparing the distribution for distances of all-atom (top row) and carbon-carbon pairs (bottom row) for reference molecules (gray) and model generated molecules (color). JSD between two distributions is reported.
• Figure 4: Conformer RMSD of the molecules sampled from different models.
• Figure 5: Visualization of chemically implausible local structures generated by TargetDiff, DecompDiff, FLAG. Incorrect bond angles are marked by yellow circles (PDBID is 2Z3H).