🧬 biology

GeoLinker: An Equivariant Diffusion Model for 3D Molecular Linker Design

GeoLinker is a unified SE(3)-equivariant diffusion framework that addresses the challenge of 3D molecular linker design in fragment-based drug discovery by supporting multiple conditioning scenarios through anchor-aware attention and size embeddings, achieving high novelty and physical plausibility across diverse benchmarks.

Original authors: Srajal Tiwari, Dolly Sharma, Ranjit Kumar

Published 2026-09-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Srajal Tiwari, Dolly Sharma, Ranjit Kumar

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the vast landscape of drug discovery, scientists often face a puzzle that is less about finding a single key and more about crafting a bridge. The strategy, known as fragment-based drug discovery, involves taking small, simple pieces of molecules that are known to stick weakly to a disease target and trying to join them together. Imagine two separate islands that a ship can barely touch; the goal is to build a sturdy bridge between them so the ship can travel freely. This bridge is called a linker. For the drug to work, this bridge must not only connect the two pieces but also hold them in a very specific three-dimensional shape, ensuring they fit perfectly into the biological machinery they are meant to stop. Creating these bridges is difficult because the space of possible shapes is enormous, and a bridge that looks good on paper might be impossible to build in a lab or might collapse under its own weight.

Researchers at the Shiv Nadar Institution of Eminence have developed a new tool called GeoLinker to solve this problem. Instead of relying on old methods that build molecules piece by piece or search through massive databases of existing chemicals, GeoLinker uses a type of artificial intelligence that learns to imagine the shape of a bridge from scratch. The system is built on a concept called diffusion, which can be thought of as a process of starting with a cloud of random noise and slowly refining it until a clear, usable structure emerges. What makes GeoLinker special is its ability to understand the rules of three-dimensional space. It knows that if you rotate a molecule, its chemical properties do not change, and it uses this knowledge to generate bridges that are physically realistic and chemically valid.

The researchers trained this system on a massive collection of known drug-like molecules to learn how atoms naturally connect. Once trained, they tested GeoLinker in four different ways to see how flexible it could be. Sometimes they asked it to build a bridge with no specific instructions, other times they told it exactly which atoms to attach to, and in other cases, they specified how long the bridge should be. They also tested a mode where it had to follow both the attachment points and the length requirements at the same time. The results showed that GeoLinker could generate a wide variety of new structures that had never been seen before. On a standard test set of molecules, the system produced new designs in nearly 80% of cases, a rate of novelty that was significantly higher than previous methods. While older tools tended to produce safe, predictable, but repetitive designs, GeoLinker explored a much wider range of shapes, creating molecules with more complex ring structures and diverse connections.

However, this exploration of new shapes came with a trade-off. Because GeoLinker prioritizes finding unique and complex structures, the percentage of molecules it created that were perfectly chemically valid was lower than that of older, more conservative methods. About 45% of the molecules generated by GeoLinker passed strict chemical checks, compared to over 90% for some previous tools. The researchers acknowledge this difference, noting that their approach is designed to cast a wider net to find rare and promising candidates, even if it means sifting through more imperfect designs. The system also proved its worth when tested on more complex, real-world scenarios involving proteins found in the human body. Even without being explicitly taught about these specific proteins, GeoLinker successfully generated bridges that connected fragments in ways that resembled known, effective drugs. In one specific case involving a protein called Hsp90, the system managed to recreate the shape of a known inhibitor, connecting two fragments with a three-carbon chain that matched the geometry of a successful drug. In another case involving a different protein, it generated a molecule that looked very similar to a known potent inhibitor, preserving the essential parts of the original fragments while building a new bridge between them.

The study suggests that this unified approach, which can handle different types of design constraints within a single framework, offers a powerful new way to think about drug design. By allowing scientists to switch between different levels of guidance—from total freedom to strict constraints—GeoLinker provides a flexible tool for exploring the chemical space. The researchers point out that while the current system is excellent at generating diverse shapes, the next step is to improve the chemical validity of the results, perhaps by teaching the system to predict how atoms bond together as it builds the shape. For now, GeoLinker stands as a significant step forward, proving that artificial intelligence can be trained to navigate the complex, three-dimensional world of molecules, offering a fresh perspective on how to build the bridges that might one day become life-saving medicines.

Technical Summary: GeoLinker

Problem Statement

Fragment-Based Drug Discovery (FBDD) is a systematic strategy for assembling high-affinity leads from low-molecular-weight fragments. A critical bottleneck in this process is molecular linker design: the task of connecting two or more fragments with a chemically valid, synthetically feasible linker that preserves the original binding geometry of the fragments.

Existing generative methods face two primary limitations:

Representation Constraints: Many approaches rely on 1D/2D representations (e.g., SMILES) or sequential graph construction, which struggle to capture the complex 3D geometric constraints required for preserving binding modes.
Lack of Flexibility: Current state-of-the-art equivariant diffusion models (e.g., DiffLinker) are typically optimized for specific conditioning settings. Practical drug discovery workflows often require switching between different constraint configurations (e.g., unconstrained generation vs. anchor-guided vs. size-conditioned) within the same project, necessitating multiple models or retraining.

Methodology

The authors propose GeoLinker, a unified SE(3)-equivariant diffusion framework designed to handle multiple fragment-linking scenarios within a single training architecture.

Core Architecture

Backbone: The model utilizes an Equivariant Graph Neural Network (EGNN) with six layers. It processes invariant scalar features (atom types) and equivariant coordinate features (3D positions).
SE(3)-Equivariance: The network ensures that the generated structures transform consistently with global rotations and translations. This is achieved by constructing edge messages from scalar node features and squared pairwise distances, updating coordinates via weighted relative displacement vectors.
Diffusion Process: GeoLinker employs a Denoising Diffusion Probabilistic Model (DDPM) formulated as a conditional geometric inpainting problem.
- Forward Process: Noise is added to linker atom positions over 1,000 timesteps while fragment coordinates remain fixed.
- Reverse Process: The model predicts clean coordinates ( $x_0$ ) from noisy inputs. Fragment positions are replaced with ground-truth coordinates at every denoising step to enforce geometric consistency.

Unified Conditioning Framework

A single model supports four distinct design modes via dynamic masking and specific conditioning modules:

Base (Unconstrained): Generates linkers conditioned only on fragment geometry.
Anchor-Guided: Incorporates known attachment atoms using anchor-aware cross-attention, where linker representations attend to anchor representations. Anchor positions are held fixed during reverse diffusion.
Size-Conditioned: Conditions generation on a target linker atom count ( $N_{link}$ ) using learned size embeddings projected into the node feature space.
Joint (Both): Combines anchor cross-attention and size conditioning for fully constrained generation.

Training Objectives and Losses

The model is trained on the ZINC dataset (438,610 molecules) using a unified loss function that balances coordinate reconstruction with geometric plausibility:

Denoising Loss ( $L_{denoise}$ ): Mean Squared Error (MSE) between predicted and ground-truth linker coordinates.
Atom Type Loss ( $L_{type}$ ): Cross-entropy loss for atom classification.
Auxiliary Geometric Losses:
- Endpoint Distance Loss: Penalizes the distance between fragment anchors and the nearest generated linker atoms, targeting a 1.5 Å bond length.
- Non-Anchor Repulsion Loss: Prevents linker atoms from forming bonds with non-anchor fragment atoms (within 1.8 Å).
- Steric Clash Loss: Applies a quadratic penalty for inter-atomic distances below 1.2 Å to prevent unphysical overlaps.

Post-Processing

Generated 3D point clouds are converted into molecular graphs using a greedy valence-constrained perception procedure followed by MMFF94 relaxation and RDKit sanitization.

Key Results

GeoLinker was evaluated on three benchmarks: ZINC (in-distribution), CASF-2016 (bioactive protein-ligand complexes), and GEOM-drugs (structurally diverse conformers).

Diversity vs. Validity Trade-off:
- On the ZINC benchmark, GeoLinker achieved significantly higher Uniqueness (86.4%–88.7%) and Novelty (78.7%–82.1%) compared to baselines like DiffLinker (Uniqueness: 22.7%–51.4%; Novelty: 30.3%–47.7%).
- GeoLinker generated linkers with higher topological complexity (0.54–0.63 rings per linker vs. 0.21–0.36 for baselines).
- However, this came at the cost of lower Chemical Validity (44.1%–45.5%) compared to graph-based methods (up to 99.3%). The authors attribute this to the coordinate-first approach exploring a broader structural space, which sometimes results in topologies that fail strict sanitization.
Generalization:
- CASF-2016: GeoLinker demonstrated zero-shot generalization to bioactive poses without protein-pocket conditioning, achieving ~58% validity and ~71% novelty.
- GEOM-drugs: The model maintained high diversity (98%+ uniqueness/novelty) across structurally heterogeneous molecules.
Case Studies:
- Hsp90: GeoLinker successfully reconstructed a three-carbon linker topology connecting two fragments in the Hsp90 ATPase pocket, recovering the topology of a known inhibitor (Tanimoto similarity = 0.32).
- IMPDH: The model generated a molecule connecting complex fragments with a Tanimoto similarity of 0.643 to a reference inhibitor, preserving input fragment geometry.

Significance and Claims

The paper positions GeoLinker as a unified framework that addresses the practical need for flexibility in fragment-based design. Its primary contributions are:

Single-Model Versatility: It supports four distinct constraint regimes (unconstrained, anchor-guided, size-conditioned, and joint) within a single SE(3)-equivariant architecture, eliminating the need for separate models for different workflow stages.
Enhanced Structural Diversity: By prioritizing 3D geometric generation over discrete graph topology, GeoLinker explores a broader chemical space, yielding significantly higher novelty and ring complexity than existing diffusion baselines.
Practical Applicability: The framework demonstrates the ability to generate plausible linkers for bioactive targets (CASF-2016) and diverse chemical spaces (GEOM) without explicit protein-pocket conditioning.

The authors explicitly acknowledge a diversity-validity trade-off, noting that while GeoLinker excels in exploring diverse scaffolds, its chemical validity is lower than graph-based approaches. They frame this as a feature for early-stage discovery where exploring novel bridging scaffolds is prioritized, while identifying the integration of direct bond prediction and synthesis constraints as necessary future work to improve validity.