Structure-aware geometric graph learning for modeling protease-substrate specificity at scale

The paper introduces OmniCleave, a scalable, structure-aware geometric graph learning framework that outperforms existing methods in modeling protease-substrate specificity by integrating multi-scale structural graphs and higher-order relational topology, thereby enabling the discovery of novel substrates and cleavage sites across diverse protease families.

The Gist

Imagine your body is a bustling city, and proteases are the specialized construction crews (or scissors) that constantly cut, trim, and reshape the buildings (proteins) to keep the city running smoothly. Sometimes they fix a broken bridge; other times, they demolish an old building to make room for a new one.

But here's the problem: These crews are picky. A specific crew only cuts a building at a very specific spot, and only if the building has the right shape and neighborhood context. If they cut in the wrong place, the city could collapse, leading to diseases like cancer or Alzheimer's.

For a long time, scientists tried to predict where these crews would cut using a simple rulebook: "If the building has these 4 letters in a row, cut here." This is like trying to guess where a tailor will cut a piece of fabric just by looking at the pattern on the surface, ignoring how the fabric folds, stretches, or what's underneath it. It works okay, but it misses the big picture.

Enter OmniCleave: The "3D City Planner"

The researchers behind this paper built a new, super-smart AI tool called OmniCleave. Instead of just reading the "letter pattern" (the sequence), OmniCleave looks at the 3D structure of the building and the relationships between the construction crews.

Here is how it works, broken down with simple analogies:

1. The "Micro-Neighborhood" View (Structure-Aware)

Imagine you are trying to guess where a gardener will prune a tree.

• Old Method: You just look at the leaves on the branch. "Oh, this leaf is green, so maybe cut here."
• OmniCleave: It zooms in and looks at the entire micro-neighborhood around the branch. It sees how the branch twists, how close the neighboring branches are, and the "energy" of the wood. It builds a hierarchical map:
  • Residue Level: It looks at the "rooms" (amino acids) in the building.
  • Atomic Level: It looks at the "bricks and mortar" (individual atoms) inside those rooms.
  • Why it matters: Just like a cut might be easy in a loose knot of rope but impossible in a tight, knotted section, OmniCleave understands the physical shape and tension of the protein to know exactly where the scissors can fit.

2. The "Crew Network" (Protease-Protease Interaction)

In the old days, scientists studied each construction crew in isolation. "Crew A cuts here. Crew B cuts there."

• The Problem: In reality, crews talk to each other. If Crew A is busy, Crew B might step in. Sometimes, two crews work together to take down a building.
• OmniCleave's Solution: It maps out a social network of all the protease crews. It knows that "Caspase-3" and "Caspase-7" are best friends and often do similar jobs. By understanding these relationships, if OmniCleave sees a building that looks like a target for Crew A, it can use its knowledge of Crew B to make a better guess, even if it hasn't seen that specific building before. It's like a detective who solves a crime not just by looking at the suspect, but by knowing the suspect's whole gang.

3. The "Universal Translator" (Geometric Graph Learning)

OmniCleave uses a special type of math called Geometric Graph Learning.

• Think of the protein as a Lego structure.
• OmniCleave doesn't just look at the color of the Legos (the sequence); it looks at how they are snapped together in 3D space.
• It creates a "universal map" where it can learn from one type of crew and apply that knowledge to a completely different type of crew. This allows it to scale up and predict cuts for over 100 different types of proteases at once, something previous tools couldn't do well.

The Proof: Did it work?

The researchers didn't just build the tool; they tested it in the real world.

• The Benchmark: They compared OmniCleave against six other top tools. OmniCleave won almost every time, especially when the "buildings" were complex and the "crews" were working together.
• The Real-World Test: They picked three new proteins that OmniCleave predicted would be cut by Caspase-3 (a crew involved in cell death). They went into a lab, mixed the proteins with the enzyme, and watched.
  • Result: The proteins did get cut exactly where OmniCleave predicted!
  • Bonus: They found new cutting spots that older tools missed. It's like finding a hidden door in a building that everyone thought was solid wall.

Why Should You Care?

This isn't just about math; it's about medicine.

• Drug Discovery: If we know exactly where these "scissors" cut, we can design better drugs to stop them from cutting the wrong things (which causes disease) or to help them cut the right things (to cure disease).
• Personalized Medicine: It helps us understand why a specific person's body might be breaking down proteins too fast or too slow.

In a nutshell:
OmniCleave is like upgrading from a 2D paper map to a 3D, real-time GPS that also knows the social lives of the drivers. It doesn't just guess where the cut will happen; it understands the shape of the road, the traffic, and the drivers' habits to predict the future with incredible accuracy.

Technical Summary

1. Problem Statement

Proteases are enzymes critical for cellular regulation, signaling, and disease pathogenesis. Accurately predicting protease-substrate cleavage sites is essential for understanding biological processes but remains a significant computational challenge.

• Limitations of Current Methods: Existing computational tools primarily rely on sequence-based representations (e.g., Position-Specific Scoring Matrices, simple machine learning on amino acid sequences). These approaches fail to capture the spatial constraints, 3D structural context, and higher-order relationships (such as protease-protease interactions) that govern enzymatic specificity.
• Scalability and Complexity: Most current models are protease-specific (trained on one protease at a time), failing to model the complex "many-to-one" relationships where multiple proteases target the same substrate or where protease interactions influence substrate recognition.
• Data Gap: There is a lack of unified frameworks that can systematically learn protease-substrate interactions across diverse protease families using structural data at scale.

2. Methodology: OmniCleave Framework

The authors introduce OmniCleave, a novel deep learning framework that integrates structure-aware geometric graph learning with protease-protease interaction networks to predict cleavage sites across 103 proteases in a single inference step.

A. Data Curation

• Dataset: 57,278 protease-substrate pairs derived from MEROPS and UniProt, covering 9,651 substrates and 103 proteases across six superfamilies (Aspartic, Cysteine, Glutamic, Metallopeptidase, Asparagine, and Serine).
• Structural Data: 3D structures were retrieved from AlphaFold DB or generated via ESMFold.
• Balancing: An equal number of non-cleavage sites were randomly sampled to create a balanced binary classification task.

B. Core Architecture

OmniCleave consists of three primary modules:

1. Cleavage-Centric Hierarchical Encoder:

   • Constructs local subgraphs centered on the cleavage site (P1 residue) within a 10 Å radius.
   • Hierarchical Representation:
     • Residue Level: Nodes represent residues with features including ESM-2 embeddings (sequence context), DSSP annotations (secondary structure), and Rosetta energy terms (thermodynamic stability).
     • Atom Level: Nodes represent individual atoms with learnable embeddings for atom type and spatial position.
   • Graph Neural Network: Uses a Generalist Equivariant Transformer (GET) to update features across both atom and residue levels, ensuring E(3)-equivariance (rotation/translation invariance). This captures local geometric context and distal residue interactions.

2. Protease-Protease Interaction (PPI) Module:

   • Constructs a graph where proteases are nodes and edges represent interactions derived from the STRING database.
   • Uses a Graph Neural Network (GNN) to encode relational priors, allowing the model to learn cooperative patterns and evolutionary relationships between different proteases. This enables the model to generalize across protease families.

3. Graph Transformer-Based Interaction Module:

   • Integrates the protease PPI graph with the substrate cleavage-site subgraphs into a unified heterogeneous graph.
   • Employs Graph Transformer Convolutions (GTC) to propagate messages between proteases and substrate sites.
   • Output: A link prediction task that outputs a scalar probability for the likelihood of a specific protease cleaving a specific site.

3. Key Contributions

• Unified Scalable Framework: Unlike previous tools that require training separate models for each protease, OmniCleave predicts sites for 103 proteases simultaneously, handling complex "many-to-one" interactions effectively.
• Structure-Aware Geometric Learning: It moves beyond sequence motifs by explicitly encoding 3D spatial arrangements, atomic distances, and energy landscapes using hierarchical geometric graphs.
• Integration of PPI Networks: It is the first method to leverage protease-protease interaction networks as prior knowledge to improve substrate specificity prediction, capturing cellular-level regulatory contexts.
• Interpretability: The model provides mechanistic insights by identifying which structural features (e.g., specific secondary structures, energy terms, distal residues) drive predictions.

4. Results

• Benchmark Performance:
  • OmniCleave outperformed six state-of-the-art (SOTA) methods (including DeepCleave, PROSPERous, and Procleave) across 103 proteases.
  • Achieved AUC > 0.9 for 48 proteases and AUC > 0.8 for 75 proteases.
  • Maintained high performance even under strict protein similarity thresholds (<30%), demonstrating strong generalization.
• Many-to-One Interaction Handling:
  • In scenarios where a single cleavage site is targeted by multiple proteases (up to 20), OmniCleave maintained near-complete coverage, whereas other methods' performance dropped significantly.
• Structural Context Validation:
  • Secondary Structure: Predictions accurately mirrored native distributions (e.g., cleavage sites frequently located in loops and $\alpha$-helices).
  • Distal Residues: Case studies (e.g., Cathepsin L) showed OmniCleave could identify cleavage sites influenced by residues far from the immediate sequence window, which sequence-based tools missed.
  • Comparison with AlphaFold3: While AlphaFold3 is a structure predictor, OmniCleave significantly outperformed it in identifying specific cleavage sites within protein complexes, correctly identifying all annotated sites in several test cases where AlphaFold3 failed.
• Experimental Validation:
  • Novel Predictions: The model predicted three novel Caspase-3 substrates: THOC5, RPIA, and CUL7.
  • In Vitro Assays: SDS-PAGE and LC-MS/MS confirmed proteolytic cleavage for all three proteins.
  • Accuracy: OmniCleave correctly predicted 8/12, 5/6, and 8/12 cleavage sites for THOC5, RPIA, and CUL7, respectively, significantly outperforming Procleave (which predicted 0, 2, and 2 sites).
  • Mechanistic Insight: Molecular docking revealed that the predicted sites involved canonical catalytic triad interactions (hydrogen bonds, salt bridges) consistent with Caspase-3 mechanisms.

5. Significance

• Biological Discovery: OmniCleave provides a scalable tool for systematic analysis of protease biology, uncovering novel substrates and functional roles (e.g., Caspase-3's role in synaptic function and transport).
• Therapeutic Applications: The framework aids in the design of multi-targeted small-molecule inhibitors and the de novo design of proteases with custom catalytic properties by linking structural determinants to specificity.
• Methodological Advancement: It establishes a new paradigm for modeling enzyme specificity by unifying sequence, 3D structure, and protein-protein interaction networks within a geometric deep learning framework, bridging the gap between statistical correlation and structural mechanism.

In summary, OmniCleave represents a significant leap forward in computational proteomics, offering a highly accurate, interpretable, and scalable solution for modeling the complex structural and relational determinants of protease-substrate specificity.