Research map / Probabilistic and scientific modeling / Molecules and proteins

Binding-Adaptive Diffusion Models for Structure-Based Drug Design

AAAI 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based molecule design with diffusion · 11 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BindDM, a single-stage SE(3)-equivariant diffusion model for structure-based drug design that adaptively extracts a binding subcomplex at each denoising step and fuses it with the full complex via cross-hierarchy interaction nodes.

Problem

Structure-based drug design needs 3D ligand generation that captures essential protein-ligand interactions, but existing diffusion and autoregressive methods ignore binding-specific substructures of the complex.

Approach

A learnable structural pooling selects the top-ranked half of complex nodes at each denoising step to form a subcomplex; SE(3)-equivariant GNNs process the complex and subcomplex hierarchically, exchanging information through cross-hierarchy interaction nodes with gated GRU updates.

Key results

• BindDM achieves -5.92 Avg. Vina Score, versus -5.67 for DecompDiff and -5.14 for Pocket2Mol.
• BindDM reaches 64.8% average High Affinity, above DecompDiff's 64.4%, and Vina Dock Avg. -8.41 vs DecompDiff -8.39.
• Ablation removing the cross-hierarchy interaction nodes or using random subcomplex selection degrades Vina Score (e.g., Exp1 random selection -4.79 vs -5.92).

Measured effects reported in the paper

Surprising findings

• Random subcomplex selection (Exp1) yields a worse Vina Score than the full model, but its median Vina Score (-5.92) is comparable to BindDM's, suggesting the average and median metrics diverge.
• Inference cost increases substantially: BindDM takes 3372 s per 100 molecules at r=0.5 versus 1729 s for the baseline.

Limitations

• Inference time is higher than the baseline diffusion model and several competitors.
• QED and SA are weaker than Pocket2Mol and AR on some metrics; the paper itself de-emphasizes these metrics.
• Subcomplex prediction accuracy (around 0.65) is only moderately above random (around 0.5) relative to BAPNet references.

Experimental setup

Datasets: CrossDocked2020

Benchmarks: CrossDocked2020

Metrics: Vina Score Vina Min Vina Dock High Affinity QED SA Diversity JSD bond distance

Baselines: LiGAN GraphBP AR Pocket2Mol TargetDiff DecompDiff

Builds on: TargetDiff DDPM E(n)-equivariant GNN

Closest papers by meaning

• CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.88)
• Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.88)
• BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.88)
• Interaction-based Retrieval-augmented Diffusion Models for Protein-specific 3D Molecule Generation (ICML 2024 · similarity 0.87)
• FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling (NeurIPS 2024 · similarity 0.87)
• M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.87)