Abstract
Predicting interactions between proteins and other biomolecules purely based on structure is an unsolved problem in biology. A high-level description of protein structure, the molecular surface, displays patterns of chemical and geometric features that fingerprint a protein’s modes of interactions with other biomolecules. We hypothesize that proteins performing similar interactions may share common fingerprints, independent of their evolutionary history. Fingerprints may be difficult to grasp by visual analysis but could be learned from large-scale datasets. We present MaSIF, a conceptual framework based on a new geometric deep learning method to capture fingerprints that are important for specific biomolecular interactions. We showcase MaSIF with three prediction challenges: protein pocket-ligand prediction, protein-protein interaction site prediction, and ultrafast scanning of protein surfaces for prediction of protein-protein complexes. We anticipate that our conceptual framework will lead to improvements in our understanding of protein function and design.
Introduction
In biology, function follows form. Understanding a protein's role requires identifying its interaction partners, whether they are small-molecule ligands or other proteins. Traditionally, researchers have relied on sequence homology—the idea that evolutionarily related proteins likely share similar functions. However, this approach fails for "orphan" proteins with no known relatives or for de novo proteins designed from scratch. Furthermore, traditional docking simulations, which physically model molecular fit, are often too computationally expensive for large-scale database searches.
The authors propose that the "molecular surface" is the true biological interface. They introduce MaSIF (Molecular Surface Interaction Fingerprinting), a framework that uses geometric deep learning to automatically learn the complex patterns of chemistry and geometry on these surfaces that dictate binding.
Methods: The MaSIF Framework
MaSIF treats proteins as continuous shapes rather than just strings of amino acids or collections of atoms. The process begins by computing the "solvent-excluded surface" of a protein and discretizing it into a mesh.
Every point on this mesh is assigned five key features:
- Geometric Features: Shape index and distance-dependent curvature.
- Chemical Features: Hydropathy, continuum electrostatics, and hydrogen-bonding potential (free electrons/proton donors).
The core innovation is the use of geodesic convolutions. The surface is decomposed into overlapping radial patches. A neural network "slides" over these patches in the protein's native, curved 3D space. To ensure the resulting "fingerprint" is independent of the protein's orientation, the model performs multiple rotations of each patch and uses a max-pooling operation to achieve rotation invariance.
Key Insights and Results
By focusing on the surface, MaSIF can identify similar interaction patterns in proteins with completely different internal folds or sequences. This abstraction allows for "vectorized docking," where surface patches are converted into 80-dimensional mathematical vectors. This transforms the docking problem into a simple, ultrafast nearest-neighbor search.
Feature Importance and Ligand Prediction
A key finding from ablation studies was the relative importance of chemical versus geometric features. For instance, when distinguishing between structurally similar ligands like NAD and NADP—which differ by only a single phosphate group—chemical features like electrostatics proved far more critical than geometry.
MaSIF was benchmarked across three major challenges:
- Pocket Classification (MaSIF-ligand): Achieved 73% balanced accuracy in identifying which of seven cofactors bound to a pocket, far exceeding random chance (14%).
- Interface Site Prediction (MaSIF-site): Predicted protein-protein interaction (PPI) sites with an AUC of 0.85, significantly outperforming established tools like SPPIDER (AUC 0.65).
- Ultrafast PPI Search (MaSIF-search): Recovered 62 out of 100 protein complexes in just 17 minutes, whereas traditional tools like PatchDock took 48 hours to solve only 40.
Conclusion and Future Directions
While MaSIF excels at identifying tight interfaces with high shape complementarity, it faces challenges with "flat" or highly solvated interfaces. Another limitation is the "apo" problem: performance drops when testing on unbound proteins that undergo conformational changes upon binding.
Future work aims to address these issues by using molecular dynamics simulations to generate "synthetic" unbound data. Additionally, the authors envision integrating MaSIF into protein design software like Rosetta, allowing researchers to "paint" specific interaction fingerprints onto newly designed protein surfaces to engineer desired functions.
Summary updated