Abstract
Efficient molecular featurization is one of the major issues for machine learning models in drug design. Here we propose persistent Ricci curvature (PRC), in particular Ollivier persistent Ricci curvature (OPRC), for the molecular featurization and feature engineering, for the first time. Filtration process proposed in persistent homology is employed to generate a series of nested molecular graphs. Persistence and variation of Ollivier Ricci curvatures on these nested graphs are defined as Ollivier persistent Ricci curvature. Moreover, persistent attributes, which are statistical and combinatorial properties of OPRCs during the filtration process, are used as molecular descriptors, and further combined with machine learning models, in particular, gradient boosting tree (GBT). Our OPRC-GBT model is used in the prediction of protein-ligand binding affinity, which is one of key steps in drug design. Based on three most-commonly used datasets from the well-established protein-ligand binding databank, i.e., PDBbind, we intensively test our model and compare with existing models. It has been found that our model are better than all machine learning models with traditional molecular descriptors.
Introduction
In computer-aided drug design (CADD), the primary challenge lies in molecular featurization—the process of converting complex 3D biological structures into numerical representations suitable for machine learning. Traditional descriptors often rely on fixed-scale geometries or simple chemical properties, failing to capture the hierarchical nature of biomolecules, which involve interactions ranging from tight covalent bonds to long-range van der Waals forces. This study introduces a novel approach using discrete differential geometry to quantitatively characterize the relationship between molecular structure and biological function across multiple scales.
Methods
The proposed framework integrates discrete Ricci curvature with the concept of persistent homology to create a multiscale geometric "fingerprint."
- Graph Representation: Molecules are modeled as graphs where atoms serve as vertices and interactions, defined by Euclidean distance thresholds, serve as edges.
- Ollivier Ricci Curvature (ORC): ORC is calculated on these graphs to measure the connectivity between neighborhoods. Positive curvature identifies dense atomic clusters or communities, while negative curvature highlights "bridge" regions or flexible chains.
- Filtration and Persistence: By varying the distance threshold, a series of nested graphs is generated. This "filtration" process allows the model to track how geometric features persist or evolve across different spatial scales, capturing both local and global structural nuances.
- Feature Engineering: Ten statistical and combinatorial attributes are extracted from the ORC values across filtration steps. To enhance chemical resolution, 36 element-specific bipartite graphs (e.g., Carbon-Oxygen, Nitrogen-Nitrogen) are constructed.
Results and Discussion
The OPRC-GBT model was evaluated using three benchmark datasets from the PDBbind databank (v2007, v2013, and v2016). The model demonstrated state-of-the-art performance, achieving a Pearson Correlation Coefficient (PCC) of 0.838 and a Root Mean Square Error (RMSE) of 1.736 kcal/mol on the PDBbind-v2016 core set.
Comparisons with traditional machine learning models, such as those utilizing standard molecular fingerprints or scoring functions like AutoDock Vina and RF-Score, showed that OPRC-GBT consistently provides superior predictive accuracy. The results underscore that Ricci curvature serves as an effective mathematical proxy for physical aggregation and molecular interaction.
Conclusion
Ollivier Persistent Ricci Curvature offers a robust and descriptive framework for molecular featurization. While the current approach involves high-dimensional feature engineering (roughly 123,000 features per complex), it sets a new benchmark for binding affinity prediction. Future work will focus on integrating these geometric insights into Graph Neural Networks (GNNs) to automate feature extraction and further refine the predictive power of geometric deep learning in drug discovery.
Summary updated