Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization

TL;DR

This paper introduces AliDiff, a novel framework that aligns pretrained target-aware diffusion models with desired functional properties for drug design, improving binding affinity while maintaining molecular quality.

Contribution

AliDiff is a new alignment method that steers diffusion models towards molecules with higher binding affinity using an exact energy optimization approach.

Findings

Achieves state-of-the-art binding energies with up to -7.07 Avg. Vina Score.

Effectively aligns diffusion models with user-defined property preferences.

Maintains strong molecular properties during optimization.

Abstract

Generating ligand molecules for specific protein targets, known as structure-based drug design, is a fundamental problem in therapeutics development and biological discovery. Recently, target-aware generative models, especially diffusion models, have shown great promise in modeling protein-ligand interactions and generating candidate drugs. However, existing models primarily focus on learning the chemical distribution of all drug candidates, which lacks effective steerability on the chemical quality of model generations. In this paper, we propose a novel and general alignment framework to align pretrained target diffusion models with preferred functional properties, named AliDiff. AliDiff shifts the target-conditioned chemical distribution towards regions with higher binding affinity and structural rationality, specified by user-defined reward functions, via the preference optimization approach. To avoid the overfitting problem in common preference optimization objectives, we further develop an improved Exact Energy Preference Optimization method to yield an exact and efficient alignment of the diffusion models, and provide the closed-form expression for the converged distribution. Empirical studies on the CrossDocked2020 benchmark show that AliDiff can generate molecules with state-of-the-art binding energies with up to -7.07 Avg. Vina Score, while maintaining strong molecular properties. Code is available at https://github.com/MinkaiXu/AliDiff.