Decomposed Direct Preference Optimization for Structure-Based Drug Design
Xiwei Cheng, Xiangxin Zhou, Yuwei Yang, Yu Bao, Quanquan Gu
TL;DR
DecompDPO is a novel structure-based optimization method that enhances diffusion models for drug design by incorporating multi-granularity preferences and physics-informed energy terms, leading to improved molecule generation and optimization.
Contribution
The paper introduces DecompDPO, a new approach that decomposes optimization objectives and integrates preference pairs at multiple levels for better drug molecule design.
Findings
Achieves up to 95.2% median high affinity in experiments.
Improves success rate for molecule generation to 36.2%.
Attains 100% median high affinity and 52.1% success in molecular optimization.
Abstract
Diffusion models have achieved promising results for Structure-Based Drug Design (SBDD). Nevertheless, high-quality protein subpocket and ligand data are relatively scarce, which hinders the models' generation capabilities. Recently, Direct Preference Optimization (DPO) has emerged as a pivotal tool for aligning generative models with human preferences. In this paper, we propose DecompDPO, a structure-based optimization method aligns diffusion models with pharmaceutical needs using multi-granularity preference pairs. DecompDPO introduces decomposition into the optimization objectives and obtains preference pairs at the molecule or decomposed substructure level based on each objective's decomposability. Additionally, DecompDPO introduces a physics-informed energy term to ensure reasonable molecular conformations in the optimization results. Notably, DecompDPO can be effectively used for…
Peer Reviews
Decision·Submitted to ICLR 2025
I find the proposed idea on the local DPO for molecular fragments very interesting and original and can potentially help in specific tasks. I also like the idea about constrained optimisation and believe it can be used in diffusion models in general to imporve the quality of the molecules. Overall, applying DPO to improve specific properties of the generated molecule makes total sense and can help generate high-quality molecules in cases when the quality of the training data is not sufficiently
**Major issues:** 1. LocalDPO looks like the main contribution of this paper. However, the evaluation of this component does not prove that the proposed mechanism is crucial for sampling molecules with optimised properties. In fact, based on Table 4, it looks like the model without LocalDPO solves the task equally well (the differences in the scores are too small). I can imagine that the chosen evaluation method does not highlight the advantages of the LocalDPO and its unique feature – locality.
Thanks for the good work on exploring the DPO method in decomposed diffusion for SBDD. I found the work to have the following strengths: 1) The method achieved decent performance on vina score, vina min, and vina dock without significantly compromising on QED and SA scores compared with the baselines selected in the paper. 2) The paper is clearly written, easy to follow, and has well-designed figures. 3) The paper introduces the use of Local DPO for optimization, which goes beyond global DPO.
1. Performance claims in the paper DecompDPO did outperform all selected baselines in terms of vina score, vina min, and vina dock, as mentioned in the paper. However, the paper did not benchmark against some other existing works that demonstrate better performance in improving the conditional sampling of diffusion models for SBDD. For example, KGDiff[1], a guidance-based method, has better performance on vina score, vina min, and vina dock. Both KGDiff and DecompDPO were benchmarked on the sam
### Motivation & ideation I believe user preference-guided optimization of generated molecules is an important frontier in structure-based drug design. The presented work is a promising extension of the capabilities of recent machine learning models for this task. It builds on various ideas established in prior works on decomposed diffusion models [1], decomposed molecule optimization [2], and preference optimization for molecules [3], and is a sensible extension thereof. Applying the diffusion
### Empirical results While the conceptual arguments for the proposed techniques are strong and well-motivated, the empirical results are currently less convincing. The performance improvement compared to the base model is only moderate in most reported metrics (e.g. QED and SA scores in Table 1). Here, it would be useful to put these results into perspective somehow. The authors could, for example, discuss potential ways to increase the performance gap between base model and optimized model. I
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsComputational Drug Discovery Methods
MethodsDirect Preference Optimization · Diffusion