A 3D pocket-aware and affinity-guided diffusion model for lead optimization

TL;DR

This work tackles lead optimization by embedding explicit protein–ligand affinity guidance into a 3D pocket-aware diffusion model. Diffleop combines forward diffusion on coordinates, atom types, and bond types with an $EGNN$ denoiser and a learned affinity predictor, guiding sampling through gradients to produce higher-affinity ligands while maintaining chemical plausibility via bond diffusion and fake-atom mechanisms. Empirical results show Diffleop outperforms state-of-the-art baselines on binding-affinity metrics and preserves key drug-like properties; ablations verify the necessity of affinity guidance and bond diffusion. Overall, Diffleop provides a scalable, end-to-end framework for structure-based lead optimization that can accelerate discovery by generating high-affinity, synthetically accessible candidates inside target pockets.

Abstract

Molecular optimization, aimed at improving binding affinity or other molecular properties, is a crucial task in drug discovery that often relies on the expertise of medicinal chemists. Recently, deep learning-based 3D generative models showed promise in enhancing the efficiency of molecular optimization. However, these models often struggle to adequately consider binding affinities with protein targets during lead optimization. Herein, we propose a 3D pocket-aware and affinity-guided diffusion model, named Diffleop, to optimize molecules with enhanced binding affinity. The model explicitly incorporates the knowledge of protein-ligand binding affinity to guide the denoising sampling for molecule generation with high affinity. The comprehensive evaluations indicated that Diffleop outperforms baseline models across multiple metrics, especially in terms of binding affinity.

Figures (2)

• Figure 1: The architecture of Diffleop. (a) The framework of Diffleop. The diffusion process gradually adds predefined noise to the atom coordinates, atom types, and bond types of the ligand inside the protein pocket. The denoising process gradually reconstructs the molecules within the protein pocket from noisy ones. The protein structures ($P$) and retained groups of ligands ($R$, i.e. scaffolds or terminal fragments) are fixed during the whole diffusion and denoising process. The binding affinity between the intact molecule and protein is predicted by TANKBind. (b) $A$ detailed denoising procedure from time step $t$ to $t-1$ using equivariant graph neural network (EGNN) with affinity guidance, in which the trained affinity predictor calculates the binding affinity $\hat{A}$ and its gradients with respect to atom coordinates, atom types and bond types to guide the molecule generation toward high binding affinity.
• Figure 2: Diffleop achieves state-of-art performance on molecular optimization through scaffold decoration and linker design tasks. ‘*’ means the models are retrained on our training set.