ProMaya: a hierarchical universal Deep Learning framework for accurate and interpretable Protein-Protein interaction identification

ProMaya is a hierarchical, interpretable deep learning framework that integrates multi-scale structural, electronic, and language-model features to achieve over 95% accuracy in identifying protein-protein interactions across diverse species, significantly outperforming existing state-of-the-art tools.

The Gist

Imagine your body is a bustling, high-tech city. In this city, proteins are the workers, machines, and vehicles. But a single worker can't build a skyscraper or run a power plant alone. They need to team up. When two proteins come together to do a job, it's called a Protein-Protein Interaction (PPI).

Mapping these teams is crucial. It tells us how our bodies fight disease, digest food, and think. But finding these teams is incredibly hard.

The Problem: The "Needle in a Haystack"

Traditionally, scientists find these protein teams by doing expensive, slow lab experiments. It's like trying to find every person shaking hands in a stadium of 100,000 people by walking around and asking everyone individually. It takes forever, costs a fortune, and you often miss the quick, fleeting handshakes.

Old computer programs tried to help, but they were like blindfolded detectives.

• Some only looked at the protein's name tag (its genetic sequence). They knew the names but not what the proteins looked like or how they fit together.
• Others looked at the blueprint (the 3D shape) but missed the subtle chemical "handshakes" happening at the atomic level.
• They often missed the "wobbly" proteins that don't have a fixed shape until they grab a partner.

The Solution: Enter ProMaya

The authors of this paper built a new AI tool called ProMaya. Think of ProMaya not as a detective, but as a super-intelligent architect with X-ray vision and a crystal ball.

Here is how ProMaya works, using simple analogies:

1. The "Mass-Density Fingerprint" (The Core Idea)

Imagine two puzzle pieces. Old tools just looked at the shape of the edges. ProMaya looks at the weight and density of the material inside the puzzle piece.

• The Analogy: If you have two magnets, they don't just stick because they are round; they stick because of how their magnetic fields align. ProMaya calculates the "Local Surface Mass Density" (LSMD). It maps out exactly where the protein is "heavy" (packed tight with atoms) and "light" (loose and airy).
• Why it matters: It realizes that proteins fit together like a key in a lock, but the "key" is made of invisible weight and charge. ProMaya sees the "weight" of the atoms, allowing it to predict if two proteins will snap together perfectly.

2. The "Hierarchical Eye" (Seeing at All Scales)

Most AI tools look at a protein from just one distance. ProMaya has a zoom lens that works at four different levels simultaneously:

• Atomic Level (Microscope): It looks at individual atoms (like seeing the grain of wood).
• Residue Level (Hand Lens): It looks at amino acid building blocks (like seeing the wood planks).
• Surface Level (Drone View): It looks at the outer skin of the protein (like seeing the whole house).
• Sequence Level (The Book): It reads the genetic code to understand the protein's history and evolution.

ProMaya combines all these views. It's like trying to understand a person: you look at their face (structure), their voice (sequence), their body language (surface), and their DNA (history) all at once to know if they are a friend or a foe.

3. The "Flexible Dancer" (Handling Chaos)

Some proteins are rigid like bricks; others are floppy like wet noodles (called Intrinsically Disordered Regions). Old tools couldn't handle the "noodles."

• ProMaya's Trick: It knows that sometimes, the "noodle" protein only gets stiff after it grabs a partner. It predicts this "dance" where a floppy protein grabs a rigid one and they both lock into place.

The Results: A Superhero Performance

The team tested ProMaya on a massive dataset involving humans, mice, plants, and even viruses (like SARS-CoV-2).

• The Score: While the best previous tools got about 80% accuracy (like a B- student), ProMaya scored 95%+ (an A+).
• The "Zero-Shot" Win: They tested it on a plant (Maize) and a virus it had never seen before. It still worked perfectly. This proves ProMaya didn't just memorize answers; it learned the physics of how proteins stick together.
• The "Cold" Discovery: They used ProMaya on a rare Himalayan plant (Picrorhiza kurrooa) to see how it makes medicine. They discovered that at cold temperatures (15°C), the plant's proteins form a tight, efficient assembly line to make medicine. At warm temperatures (25°C), the assembly line falls apart. This is a discovery that would have taken years of lab work to find, but ProMaya found it in minutes.

Why This Matters

ProMaya is like giving scientists a universal translator for the language of life.

• Speed: It can predict interactions in seconds that used to take months in the lab.
• Cost: It saves millions of dollars in expensive experiments.
• Discovery: It can find interactions in rare plants or new viruses that we don't have enough samples to test physically.

In short, ProMaya is a hierarchical, multi-scale AI that understands the "physics of friendship" between proteins. It looks at the tiny atoms, the big shapes, and the genetic history to predict who will team up, helping us design better drugs and understand life at a deeper level.

Technical Summary

1. Problem Statement

Protein-protein interactions (PPIs) are fundamental to cellular function, yet experimentally mapping the "interactome" remains slow, expensive, and prone to false positives/negatives. While computational methods have advanced, existing approaches suffer from three critical limitations:

• Unimodal Limitations: Sequence-based models lack explicit 3D structural context (geometry, topology), while structure-based models often fail to capture evolutionary constraints and intrinsically disordered regions (IDRs).
• Lack of Physical Determinants: Existing models rarely explicitly model Local Surface Mass Density (LSMD), a physical property governing van der Waals forces and hydrophobic collapse, which is crucial for binding specificity.
• Poor Generalizability: Current tools struggle with cross-species prediction, transient interactions, and rare interaction types (e.g., post-translational modifications) because they rely heavily on sequence homology or memorized templates rather than learning universal physicochemical principles.

2. Methodology: The ProMaya Framework

ProMaya is a hierarchical, multi-scale Graph Transformer framework designed to integrate atomic, residue, surface, and sequence data into a unified representation.

A. Core Hypothesis

The authors hypothesize that PPIs are governed by complementary "mass-density fingerprints" on opposing protein surfaces. These fingerprints arise from tightly packed hydrophobic cores, aromatic stacking, and buried polar networks. Additionally, the framework accounts for flexible, disorder-mediated interactions common in signaling networks.

B. Multi-Modal Feature Extraction

ProMaya processes proteins through four distinct modalities:

1. Atomic-Level Geometry:
   • Constructs an atomic graph where nodes encode element identity, partial charge, and Local Surface Mass Density (LSMD).
   • LSMD is calculated as a Gaussian-smoothed atomic packing density ($\rho_i$), capturing electron density gradients at sub-Ångström resolution.
   • Edges encode Euclidean distances, polar/azimuthal angles, and torsion terms.
2. Residue-Level Representation:
   • Constructs a residue graph encoding backbone torsion angles, relative solvent accessibility (SASA), secondary structure, and Intrinsic Disorder Regions (IDR) scores (via IUPred2A).
   • Integrates evolutionary context via ProtTrans (a large protein language model) embeddings.
3. Surface Representation:
   • Samples the solvent-accessible surface as a point cloud (1,024 points) using MSMS.
   • Encodes surface normals, curvature, electrostatic potential, and interpolated LSMD.
   • Processed via a PointNet++ encoder.
4. Sequence Embeddings:
   • Utilizes pretrained ProtTrans embeddings to capture evolutionary constraints and functional signatures.

C. Architectural Design

• Hierarchical Graph Transformer (HGT): A heterogeneous graph transformer processes the four modalities (atomic, residue, surface, sequence) simultaneously. It uses type-specific and relation-specific multi-head attention to model long-range dependencies and heterogeneous interactions.
• Cross-Level Multimodal Alignment: Iterative bidirectional cross-attention aligns features across scales (e.g., atom $\to$ residue $\to$ surface $\to$ sequence). This is enforced by auxiliary losses:
  • LSMD-alignment loss.
  • Contrastive InfoNCE loss.
  • IDR-consistency loss.
• Cross-Protein Interaction Modeling: A sparse, LSMD-guided cross-attention mechanism computes interactions between two proteins ($P_A$ and $P_B$), focusing only on atoms with high LSMD ($>0.5$) to prioritize interaction-prone regions.
• Hybrid Prediction Head: The final pair embedding is processed through a neural feed-forward network and then an XGBoost ensemble classifier to output the interaction probability.

D. Interpretability

The framework employs Grad-CAM (Gradient-weighted Class Activation Mapping) to generate heatmaps at the atomic, residue, and surface levels, identifying the specific structural determinants driving the prediction.

3. Key Contributions

1. Introduction of LSMD: ProMaya is the first PPI framework to explicitly integrate Local Surface Mass Density as a primary biophysical driver, demonstrating that atomic packing density is a superior discriminative signal compared to standard geometric features.
2. Hierarchical Multi-Scale Architecture: It successfully unifies atomic resolution (Å), residue topology (nm), and global surface morphology (µm) with evolutionary sequence data, overcoming the "sequence-structure gap."
3. Universal Generalizability: The model is designed to be species-agnostic, capable of predicting interactions in non-model organisms and cross-kingdom pairs (e.g., virus-host) without relying on sequence homology.
4. Explainability: The integration of Grad-CAM provides mechanistic insights, distinguishing between rigid hydrophobic interfaces and flexible disorder-mediated contacts.

4. Results and Performance

ProMaya was benchmarked against state-of-the-art (SOTA) tools (e.g., PPI-GNN, D-SCRIPT, DeepPPI, PEPPI) across diverse datasets.

• Overall Performance: Achieved >95% average accuracy (specifically 95.7%) and an MCC of 0.908 on the internal test set, outperforming the next best competitor (D-SCRIPT at 83.4%) by >12%.
• Ablation Studies:
  • Removing the atomic encoder caused a massive 16.5% drop in accuracy, confirming LSMD and atomic geometry as the most critical features.
  • The full multi-scale integration provided a synergistic boost over single-modality models.
• Cross-Species & Cross-Kingdom Generalization:
  • SARS-CoV-2 (Virus-Human): Achieved MCC 0.92 on viral proteins absent from training (zero-shot cross-species).
  • Zea mays (Plant): Achieved MCC 0.92 on plant proteins with <30% sequence identity to training data, where sequence-based tools performed near-randomly.
  • Mus musculus (Mouse): Maintained high accuracy (MCC 0.93) with no sequence overlap.
• Homology-Held-Out Benchmark: In a strict test using new PDB structures with <30% sequence identity and no detectable templates, ProMaya maintained 94.3% accuracy, whereas template-based methods collapsed to ~36% (random chance).
• Class Imbalance: ProMaya showed robustness in predicting rare interaction types (e.g., Post-Translational Modifications), where other tools failed significantly.

5. Significance and Application

• Biological Discovery: The framework was applied to Picrorhiza kurrooa (a medicinal Himalayan plant) to reconstruct temperature-conditioned PPI networks. It successfully identified a cold-stabilized iridoid metabolon (enzyme complex) and a temperature-induced disordering of regulatory complexes (CBF1-ICE1), providing testable hypotheses for metabolic engineering.
• Universal Tool: ProMaya serves as a "universal" predictor, capable of bypassing costly experiments for species lacking experimental interactome data.
• Accessibility: The system is available as a free web server, democratizing access to high-accuracy PPI prediction.

Conclusion

ProMaya represents a paradigm shift in PPI prediction by moving from sequence-similarity or single-modality structural approaches to a physics-informed, hierarchical multi-modal deep learning framework. By explicitly modeling Local Surface Mass Density and integrating it with evolutionary and disorder signals, ProMaya achieves unprecedented accuracy and generalizability, offering a scalable solution for mapping interactomes across the tree of life.