Research map / Probabilistic and scientific modeling / Molecules and proteins

Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization

NeurIPS 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 5 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 AliDiff, which aligns pretrained target-aware ligand diffusion models (IPDiff, TargetDiff) to binding-affinity rewards via an Exact Energy Preference Optimization (E²PO) objective, evaluated on CrossDocked2020.

Problem

Target-aware structure-based drug design diffusion models learn the distribution of all training ligands equally, so they lack steerability toward high binding affinity and chemical quality; preference optimization (DPO) for this overfits to winning samples.

Approach

Adapt diffusion preference optimization to 3D ligand generation with a Bradley-Terry-style reward target, and replace DPO's log-sigmoid loss with an E²PO cross-entropy objective that weights preference and reversed-preference losses by sigma(r^w - r^l), which provably minimizes a KL to the reward-tilted distribution.

Key results

• AliDiff reaches -7.07 Avg. Vina Score (median -7.95) versus -6.42 for IPDiff on 100 test proteins, a 10.1% improvement.
• AliDiff's High Affinity rate is 73.4% vs 69.5% for IPDiff, with small drops in QED (0.50 vs 0.52) and SA (0.57 vs 0.61).
• Applying the method to TargetDiff (AliDiff-T) improves Avg. Vina Score from -5.47 to -5.81 and QED from 0.48 to 0.56.

Measured effects reported in the paper

Surprising findings

• Choosing the worst-ranked ligand as the losing sample gave the best binding affinity, while choosing the second-best sample performed poorly.
• Supervised fine-tuning on the top 50% of samples gave no improvement over the baseline model.

Limitations

• Binding affinity rewards come from AutoDock Vina, an approximate and sometimes inaccurate scoring function.
• Training is offline with fixed preference pairs; online preference learning is not explored.
• Binding affinity gains come with slight drops in QED and SA relative to IPDiff.

Experimental setup

Datasets: CrossDocked2020

Benchmarks: CrossDocked2020

Metrics: Vina Score Vina Min Vina Dock High Affinity QED SA Diversity Lipinski

Baselines: liGAN GraphBP AR Pocket2Mol TargetDiff DecompDiff IPDiff

Builds on: DPO Diffusion-DPO IPDiff TargetDiff

Closest papers by meaning

• Interaction-based Retrieval-augmented Diffusion Models for Protein-specific 3D Molecule Generation (ICML 2024 · similarity 0.89)
• Aligning Transformers with Continuous Feedback via Energy Rank Alignment (NeurIPS 2025 · similarity 0.89)
• DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · similarity 0.88)
• Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.88)
• Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design (NeurIPS 2025 · similarity 0.87)
• Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.87)