[Paper Review] DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization

Xiangxin Zhou, Xiwei Cheng+4 more|arXiv (Cornell University)|Mar 7, 2024

Analytical Chemistry and Chromatography7 citations

TL;DR

DecompOpt integrates controllable and decomposed diffusion with iterative optimization to design and optimize structure-based ligands, enabling de novo design and controllable generation (R-group design and scaffold hopping) with improved docking affinity and drug-like properties.

ABSTRACT

Recently, 3D generative models have shown promising performances in structure-based drug design by learning to generate ligands given target binding sites. However, only modeling the target-ligand distribution can hardly fulfill one of the main goals in drug discovery -- designing novel ligands with desired properties, e.g., high binding affinity, easily synthesizable, etc. This challenge becomes particularly pronounced when the target-ligand pairs used for training do not align with these desired properties. Moreover, most existing methods aim at solving extit{de novo} design task, while many generative scenarios requiring flexible controllability, such as R-group optimization and scaffold hopping, have received little attention. In this work, we propose DecompOpt, a structure-based molecular optimization method based on a controllable and decomposed diffusion model. DecompOpt presents a new generation paradigm which combines optimization with conditional diffusion models to achieve desired properties while adhering to the molecular grammar. Additionally, DecompOpt offers a unified framework covering both extit{de novo} design and controllable generation. To achieve so, ligands are decomposed into substructures which allows fine-grained control and local optimization. Experiments show that DecompOpt can efficiently generate molecules with improved properties than strong de novo baselines, and demonstrate great potential in controllable generation tasks.

Motivation & Objective

• Motivate structure-based molecular design that jointly optimizes binding affinity and drug-like properties.
• Develop a controllable diffusion framework that decomposes ligands into arms and scaffold for fine-grained control.
• Unify de novo design and controllable generation within a decomposed diffusion paradigm.
• Demonstrate optimization-driven generation that can outperform strong baselines on CrossDocked2020.

Proposed method

• Introduce a controllable and decomposed diffusion model to generate ligands conditioned on protein subpockets and reference arms.
• Decompose ligands into scaffold and arms; condition arm-level features via SE(3)-equivariant encoders on arm-pocket pairs.
• Use a diffusion-based decoder with decomposed priors and conditional features to preserve molecular grammar while enabling control.
• Iterative optimization where arms are replaced by higher-scoring candidates; integrate docking/evaluations as part of the optimization loop.
• Adopt multi-objective scoring (QED, SA, Vina Min) with Z-score normalization to guide arm-level selection across subpockets.

Experimental results

Research questions

• RQ1 Can a controllable, decomposed diffusion model generate high-affinity ligands while respecting molecular grammar?
• RQ2 Does arm-level optimization improve efficiency and diversity over molecule-level or purely generative approaches?
• RQ3 Can the framework support controllable tasks like R-group design and scaffold hopping in 3D structure-based design?
• RQ4 How does integrating optimization with generation affect binding affinity and drug-likeness metrics compared to baselines?

Key findings

• DecompOpt achieves higher affinity-related metrics and a greater success rate compared to strong baselines on CrossDocked2020.
• For de novo design, DecompOpt attains a mean Vina Dock score of -8.98 and a 52.5% average success rate, outperforming several baselines.
• Arm-level optimization outperforms molecule-level optimization in efficiency and property gains, highlighting benefits of decomposed optimization.
• Controllability enables R-group design and scaffold hopping, with scaffold hopping yielding higher validity and complete rates and promoting diversity.
• On average, DecompOpt improves QED and SA, and achieves deeper docking improvements (e.g., Vina Dock and Vina Min) relative to the DecompDiff baseline.