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.
Introduction
Fragment-Based Drug Discovery (FBDD) is a cornerstone of modern pharmaceutical research, where small chemical fragments are identified for their binding affinity to a target protein and subsequently linked to form potent drug candidates. However, designing these linkers is a complex geometric and chemical challenge. Traditional computational methods often struggle with several limitations: they are typically restricted to connecting only two fragments, require manual specification of attachment points (anchors), and often ignore the spatial constraints imposed by the protein binding pocket, leading to steric clashes.
Methodology
DiffLinker addresses these challenges using a conditional diffusion framework. Diffusion models generate data by learning to reverse a noise-injection process. In DiffLinker, the input fragments are treated as a fixed context, and the model iteratively denoises a set of linker atoms to form a cohesive molecule.
E(3)-Equivariant Architecture
The core of DiffLinker is an Equivariant Graph Neural Network (EGNN). This architecture ensures E(3)-equivariance, meaning the model's predictions are invariant to global rotations and translations of the input fragments. This is crucial for maintaining the geometric integrity of the generated 3D structures.
Linker Size Prediction and Denoising
Unlike previous models that require a predefined linker size, DiffLinker employs a separate GNN to predict the probability distribution of the number of atoms needed for the linker based on the fragments' spatial arrangement. Once a size is sampled, the model performs a "one-shot" generation of the entire linker through a continuous diffusion process. Discrete atom types are mapped into a continuous space for denoising and then projected back to discrete categories.
Experimental Setup
The model was evaluated across several benchmarks:
- ZINC & CASF: Standard datasets for assessing molecular generation and 3D conformations.
- GEOM (Multi-Fragment): A new benchmark specifically designed to test the model's ability to link three or more fragments simultaneously.
- Protein Pockets: Data from Binding MOAD was used to evaluate the model's performance in structure-based design, where the protein pocket acts as an additional geometric constraint.
Results and Discussion
DiffLinker demonstrates significant improvements over existing baselines such as DeLinker and 3DLinker.
Chemical Quality and Diversity
The model consistently produces molecules with higher Synthetic Accessibility (SA) and Drug-Likeness (QED) scores. Notably, DiffLinker generates a higher proportion of unique and novel molecules, avoiding the mode collapse often seen in autoregressive generative models.
Multi-Fragment Linking and Pocket Awareness
On the GEOM multi-fragment benchmark, DiffLinker achieved a validity rate of approximately 94%, vastly outperforming 3DLinker, which failed in nearly 84% of cases. Furthermore, when conditioned on protein pockets, the model successfully reduced atomic clashes with the protein structure. The average number of clashes dropped from 21 in unconditioned samples to 7 in pocket-conditioned samples, a level comparable to native ligands.
Conclusion
DiffLinker represents a significant advancement in molecular linker design by providing an anchor-free, multi-fragment, and pocket-aware generative framework. By leveraging E(3)-equivariant diffusion, it produces chemically valid and diverse drug candidates that respect the complex 3D constraints of biological targets. Future work will focus on integrating bond-order prediction directly into the diffusion process and accounting for protein side-chain flexibility to further enhance the realism of the generated molecules.
Summary updated