Curvature-Guided Geometric Representation for Protein-Ligand Binding Affinity Prediction

RicciBind is a novel geometric representation framework that enhances protein-ligand binding affinity prediction by integrating Ricci curvature-guided hierarchical structure learning with optimal transport-based cross-domain alignment to effectively model both local interaction tightness and globally coordinated cross-molecular interactions.

The Gist

Imagine you are trying to figure out how well two puzzle pieces fit together. In the world of drug discovery, these "puzzle pieces" are a protein (a large, complex machine in your body) and a ligand (a small molecule, often a potential medicine). The goal is to predict how tightly they will stick together, which scientists call "binding affinity." If they stick too loosely, the drug won't work; if they stick perfectly, it might cure a disease.

For a long time, computers have tried to predict this fit using math and data. However, most existing methods are like looking at a puzzle from just one angle or counting the number of pieces without understanding their shape. They miss the subtle, 3D "hugs" and "handshakes" happening between the two molecules.

This paper introduces a new computer program called RicciBind. Think of RicciBind as a master puzzle expert who doesn't just look at the pieces, but understands the curvature and shape of the space they occupy.

Here is how RicciBind works, broken down into three simple steps using everyday analogies:

1. The "Curvature" Map (Understanding the Shape)

Imagine you are walking through a forest. Some parts are flat and open, while others are dense and crowded with trees. In math, this "crowdedness" or "sparseness" is called curvature.

RicciBind uses a special mathematical tool called Ricci Curvature to map the protein and the drug.

• The Analogy: Instead of just seeing atoms as dots, RicciBind sees them as a landscape. If a group of atoms is tightly packed together (like a dense forest), the curvature is "positive." If they are spread out or disconnected (like a sparse desert), the curvature is "negative."
• Why it helps: This helps the computer understand which parts of the molecule are "tight" and important for sticking, and which parts are loose and irrelevant. It gives the computer a better sense of the molecule's true 3D shape.

2. Grouping the Neighborhoods (Clustering)

Once the computer understands the shape, it needs to make sense of the thousands of individual atoms.

• The Analogy: Imagine a huge city with millions of people. It's too hard to talk to every single person. Instead, you group them into neighborhoods. RicciBind uses the "curvature map" to decide how to group atoms. It puts atoms that are tightly connected (positive curvature) into the same "neighborhood" or cluster.
• The Result: Instead of looking at 10,000 individual atoms, the computer now looks at a few dozen "super-clusters" that represent functional parts of the molecule. This makes the problem much easier to solve while keeping the important details.

3. The "Best Match" Dance (Optimal Transport)

Now the computer has a protein made of clusters and a drug made of clusters. How do they match?

• The Analogy: Imagine you have two groups of dancers (the protein clusters and the drug clusters). You want to pair them up so they dance together perfectly. You don't just pair them randomly; you calculate the "cost" of every possible pairing to find the most efficient, harmonious dance plan. This is called Optimal Transport.
• The Magic: RicciBind uses this math to figure out exactly which protein "neighborhood" should interact with which drug "neighborhood." It ignores the parts that don't fit and highlights the specific spots where the drug and protein lock together.

What Did They Find?

The authors tested RicciBind on many different datasets (collections of known protein-drug pairs).

• Better Accuracy: It predicted how well drugs would stick to proteins more accurately than previous methods, including other advanced AI models.
• Better Generalization: Even when the computer was shown a brand-new protein it had never seen before (a "cold start" scenario), RicciBind still performed well. It didn't just memorize the data; it learned the underlying rules of how shapes fit together.
• Virtual Screening: In a test where the computer had to find the "winning" drug among thousands of "decoys" (fake drugs), RicciBind was very good at spotting the real winners quickly.

The Bottom Line

RicciBind is a new way for computers to understand drug interactions. By using curvature to understand the shape of molecules and optimal transport to match them up like a perfect dance, it creates a clearer, more accurate picture of how medicines work. This helps scientists design better drugs faster, without needing to test every single possibility in a lab.

Technical Summary

Technical Summary: Curvature-Guided Geometric Representation for Protein–Ligand Binding Affinity Prediction

Problem Statement
Protein–ligand binding affinity (PLA) prediction is a critical component of drug discovery, yet existing machine learning approaches face significant limitations. While deep learning methods have advanced the field, they often struggle to jointly characterize the local geometric organization and globally coordinated cross-molecular interactions inherent in complex binding mechanisms. Structure-free methods neglect 3D structural information, while existing structure-based methods, particularly hierarchical 3D graph neural networks (3D-GNNs), often fail to explicitly incorporate intrinsic geometric properties of molecular graphs. This omission leads to incomplete representations of molecular geometry, limiting the models' ability to capture complex cross-scale dependencies and constraining their expressiveness and interpretability.

Methodology: RicciBind
To address these limitations, the authors propose RicciBind, a geometric representation framework that integrates curvature-guided hierarchical structure learning with optimal transport (OT)-based cross-domain alignment. The model treats proteins and ligands as two distinct heterogeneous structural domains and operates through four primary stages:

1. Graph Construction: The protein–ligand complex is encoded as a 3D molecular graph $G=(V, E, R, K)$, where nodes represent atoms and edges represent covalent bonds (within molecules) and non-covalent interactions (between molecules, defined by a distance threshold of 5 Å).
2. Curvature-Aware Graph Embedding (CAGE): This module encodes the complex into atom-level representations. It incorporates Ollivier-Ricci Curvature (ORC) as a topological descriptor within the message-passing process. By calculating ORC for edges, the model captures local connectivity patterns and global structural organization. Positive curvature indicates dense, tightly coupled local structures, while negative curvature signifies structural divergence. This curvature information guides the message-passing mechanism to refine atom-level node representations.
3. Curvature-Driven Clustering (CDC): To move beyond atom-level granularity, RicciBind employs a differentiable pooling framework guided by ORC. The curvature values are transformed into a curvature-informed adjacency matrix, which adaptively assigns atoms to clusters. This process strengthens aggregation within strongly connected (positively curved) regions and suppresses coupling across weakly connected (negatively curved) areas, yielding higher-order, geometrically consistent cluster-level representations.
4. Optimal Transport-Based Cluster Matching (OTCM): This module models the interaction between protein and ligand clusters as a cross-domain alignment problem. Using entropy-regularized optimal transport, the model computes a transport plan that aligns structurally and functionally related clusters across the protein and ligand domains. This bidirectional matching identifies key interaction clusters while attenuating the influence of irrelevant ones, capturing higher-order interaction patterns.
5. Prediction: The final protein–ligand interaction representation fuses intra-domain curvature-aware embeddings with cross-domain transport features. This unified representation is passed through a prediction head (MLP) to regress the binding affinity. The model is trained end-to-end using mean-squared error (MSE) loss.

Key Contributions
The paper highlights four primary contributions:

• Curvature-Aware Embedding: The design of a graph embedding module that integrates ORC as a topological descriptor to enhance atom-level node representations, capturing intrinsic geometric structures.
• Curvature-Driven Clustering: The proposal of a clustering module that adaptively assigns atoms to clusters under Ricci curvature guidance, producing higher-order representations consistent with the geometric–topological structure of molecular graphs.
• OT-Based Heterogeneous Modeling: To the authors' knowledge, this is the first study to treat proteins and ligands as distinct heterogeneous structural domains and model their binding complexes using optimal transport principles, providing theoretical support for aligning key functional interaction clusters.
• Empirical Superiority: Comprehensive evaluation across multiple public datasets demonstrating superior predictive performance and generalization compared to existing structure-free and structure-based approaches.

Experimental Results
RicciBind was evaluated on several benchmarks:

• Cross-Dataset PLA Prediction: On PDBbind v2013, v2016, and v2019 holdout sets, RicciBind achieved state-of-the-art or highly competitive performance, outperforming both structure-free models (e.g., DeepDTA, GraphDTA) and structure-based baselines (e.g., EHIGN, CheapNet). Notably, it achieved a Pearson correlation of 0.872 on the 2013 core set and 0.667 on the challenging 2019 holdout set.
• Generalization to Unseen Targets: In evaluations on the CSAR NRC-HiQ dataset and the Holoprot benchmark (diverse protein targets), RicciBind demonstrated strong generalization, achieving the highest Pearson correlation (0.92) on CSAR and outperforming baselines on unseen proteins.
• Cold Start Scenarios: Under rigorous cold start settings (protein sequence identity <30% and novel ligand scaffolds), RicciBind maintained leading performance, showing robustness to structural variations in both proteins and ligands.
• Virtual Screening: In structure-based virtual screening (SBVS) tasks on DUD-E (cross-family zero-shot) and LIT-PCBA (target-specific), RicciBind demonstrated superior early enrichment capabilities, significantly outperforming docking-based scoring functions and other learning-based methods.
• Ablation Studies: Removing ORC information from either the embedding or clustering modules resulted in performance degradation, confirming the essential role of Ricci curvature in enhancing molecular interaction representations.

Significance and Claims
The paper claims that RicciBind offers a promising new paradigm for structure-based drug discovery by effectively modeling the complex geometric and topological structures underlying molecular interactions. By coupling curvature-guided structure encoding with OT-driven cross-domain alignment, the framework substantially improves both the accuracy and interpretability of binding affinity prediction. The authors assert that their approach provides a principled geometric foundation for capturing higher-order interaction patterns that are often missed by conventional feature aggregation methods. While the current focus is on PLA prediction, the authors note that the underlying modeling paradigm is extensible to broader biomolecular representations and related tasks such as pocket detection and molecular docking.