[Paper Review] Equivariant 3D-Conditional Diffusion Models for Molecular Linker Design

Ilia Igashov, H. Stärk+6 more|arXiv (Cornell University)|Oct 11, 2022

Protein Structure and Dynamics35 citations

TL;DR

We introduce DiffLinker, an E(3)-equivariant 3D-conditional diffusion model that can generate linker atoms to connect an arbitrary number of molecular fragments, optionally conditioned on protein pockets, without needing predefined attachment points.

ABSTRACT

Fragment-based drug discovery has been an effective paradigm in early-stage drug development. An open challenge in this area is designing linkers between disconnected molecular fragments of interest to obtain chemically-relevant candidate drug molecules. In this work, we propose DiffLinker, an E(3)-equivariant 3D-conditional diffusion model for molecular linker design. Given a set of disconnected fragments, our model places missing atoms in between and designs a molecule incorporating all the initial fragments. Unlike previous approaches that are only able to connect pairs of molecular fragments, our method can link an arbitrary number of fragments. Additionally, the model automatically determines the number of atoms in the linker and its attachment points to the input fragments. We demonstrate that DiffLinker outperforms other methods on the standard datasets generating more diverse and synthetically-accessible molecules. Besides, we experimentally test our method in real-world applications, showing that it can successfully generate valid linkers conditioned on target protein pockets.

Motivation & Objective

• Address fragmentation-based drug design challenges by enabling linker generation between an arbitrary number of fragments.
• Eliminate the need for predefined attachment points by learning linker connections from data.
• Incorporate 3D pocket information to produce chemically valid, pocket-compatible linkers.
• Achieve state-of-the-art chemical relevance, synthetic accessibility, and diversity in linker generation.
• Demonstrate applicability to real-world structure-based drug design scenarios with pocket conditioning.

Proposed method

• Propose DiffLinker, a conditional diffusion model operating on 3D atomic point clouds with context u consisting of input fragments and optional protein pockets.
• Use an E(3)-equivariant graph neural network (EGNN) as the learnable denoising function φ, predicting noise to denoise linker coordinates and atom types.
• Define equivariance: if f is O(3)-equivariant and φ is equivariant to joint transformations of z_t and u, then p(z_0|u) is O(3)-equivariant.
• Jointly model z_t and fixed context u in a single graph where linker nodes are updated while context nodes stay fixed, ensuring translation invariance by centering on the context COM.
• Predict linker size with a separately trained GNN and sample the linker size from a learned distribution.
• Optionally condition on protein pockets by including pocket atoms in u and applying a 4 Å distance cutoff to reduce graph density, allowing clash-aware generation.

Experimental results

Research questions

• RQ1 Can an E(3)-equivariant diffusion model generate linkers for an arbitrary number of input fragments without explicit attachment point specification?
• RQ2 Does conditioning on pocket information reduce clashes and improve drug-design-relevant metrics (QED, SA, etc.) for linker generation?
• RQ3 How does DiffLinker perform relative to existing methods (DeLinker, 3DLinker) on standard benchmarks when linking more than two fragments?
• RQ4 Can the model selectively sample linker sizes to improve novelty and diversity without sacrificing key performance metrics?
• RQ5 Is the approach practical for structure-based drug design applications with realistic pocket constraints?

Key findings

• DiffLinker achieves higher QED and competitive SA and ring counts compared to baseline methods on ZINC and CASF benchmarks.
• DiffLinker yields high validity and notable diversity, with improved SC_RDKit similarity to ground-truth linkers.
• Sampling linker size improves novelty and uniqueness with minor impact on core metrics.
• When extended to GEOM with more than two fragments, DiffLinker substantially outperforms adapted 3DLinker in validity and recovery, while generating more ring-rich linkers.
• Pocket-conditioned DiffLinker can generate linkers with minimal clashes against target pockets, especially when full-atomic pocket information is provided.
• DiffLinker supports conditioning on pockets to produce realistic, pocket-aware linkers suitable for structure-based drug design.