ExPath: Targeted Pathway Inference for Biological Knowledge Bases via Graph Learning and Explanation

ExPath is a novel subgraph inference framework that integrates experimental molecular data and biological foundation models to identify biologically meaningful targeted pathways in knowledge bases, significantly outperforming existing explainers in fidelity and pathway length preservation.

The Gist

Imagine you have a massive, ancient library called the "Biological Knowledge Base." This library contains every possible road map (pathway) that exists inside a human body, showing how genes, proteins, and chemicals talk to each other. It's like a giant, static city map showing every single street, alley, and highway in the world.

The Problem:
Now, imagine you are a detective trying to solve a specific crime (a disease or a specific biological reaction) in a specific neighborhood. You have a piece of evidence (experimental data, like a specific DNA mutation).

If you just look at the giant library map, you see everything. But you don't know which specific streets are actually being used right now for this specific crime. The map is too crowded with irrelevant roads. Finding the exact route the "criminal" (the disease mechanism) took is like finding a needle in a haystack, and it usually requires a super-expert to guess.

The Solution: EXPATH
The authors of this paper built a new tool called EXPATH. Think of it as a Smart GPS with a "Highlighter" feature.

Here is how it works, broken down into simple steps:

1. The Brain: PATHMAMBA (The Detective)

First, the system needs to understand the data.

• The Old Way: Previous tools were like looking at a map and just counting how many roads connect to a specific intersection. They missed the big picture.
• The New Way (PATHMAMBA): This is a super-smart AI detective. It doesn't just look at one street at a time; it looks at the whole neighborhood and how traffic flows through the whole city.
  • The Analogy: Imagine a detective who can see the local traffic jams (local interactions) and the long-distance highway patterns (global pathways) simultaneously. It uses a special "state-space" engine (called Mamba) that is great at remembering long chains of events, like remembering a story from the beginning to the end without forgetting the middle.
  • The Result: It can look at your specific DNA evidence and say, "Ah, this specific pattern of mutations usually leads to this specific type of disease."

2. The Highlighter: PATHEXPLAINER (The Map Redrawer)

Once the detective knows what is happening, it needs to show you where it's happening.

• The Old Way: Other tools would highlight random dots on the map or individual streets. They might say, "This protein is important," and "That gene is important," but they wouldn't show you the connection between them.
• The New Way (PATHEXPLAINER): This tool acts like a digital highlighter that draws a continuous, glowing line through the map. It finds the exact chain of events (the pathway) that is responsible for the result.
  • The Analogy: Instead of just pointing at a few houses, it draws a red line from the suspect's house, through the alley, to the getaway car. It filters out all the noise and only shows the essential route.

3. The "Why" (The Explanation)

The best part is that this isn't a "black box." The system explains why it chose that path.

• It asks: "If I remove this specific road, does the detective still know the answer?"
• If the answer is "No," then that road is essential.
• If the answer is "Yes," then that road was just noise.
• This ensures the system only highlights the roads that truly matter for the specific disease or condition you are studying.

Why is this a big deal?

The researchers tested this on 301 different biological maps (covering diseases, metabolism, and cell processes).

• Accuracy: It was 4.5 times better at finding the "necessary" parts of the map (the roads you must have) compared to older tools.
• Efficiency: It was 14 times better at ignoring the "unnecessary" parts (the roads you can ignore).
• Length: It could trace paths that were 4 times longer than other methods, meaning it could find complex, multi-step chains of events that others missed.

The Real-World Test: The T-Cell Case Study

To prove it worked, they tested it on the T-Cell Receptor, which is like the "alarm system" of your immune system.

• Old Tools: Drew a messy, scattered map with red dots everywhere, making it hard to see the actual alarm signal.
• EXPATH: Drew a clean, continuous red line showing exactly how the alarm travels from the outside of the cell to the nucleus, highlighting the specific "PI3K-AKT" and "NF-κB" highways that control cell survival.

In Summary

EXPATH is like upgrading from a dusty, static paper map to a dynamic, intelligent GPS. It doesn't just show you the whole city; it looks at your specific destination (the disease), checks your current traffic (your experimental data), and draws the shortest, most accurate, and most meaningful route for you to follow, explaining exactly why that route was chosen. This helps scientists understand diseases faster and design better treatments without getting lost in the noise.

Technical Summary

1. Problem Definition

Biological knowledge bases (e.g., KEGG, STRING) contain comprehensive, static maps of molecular interactions (genes, proteins). However, these general networks lack specificity for experimental data (e.g., specific mutations, gene expression changes, or amino acid sequences).

• The Challenge: Researchers need to identify which specific sub-networks (pathways) are "active" or "targeted" under specific experimental conditions to understand disease mechanisms or cellular responses.
• Limitations of Existing Methods:
  • Statistical/Topology-driven methods: Rely on node centrality or degree but cannot incorporate experimental data features.
  • Standard GNNs: Often treat all interactions equally, failing to capture long-range dependencies inherent in biological pathways (multi-step signaling).
  • Explainability: Existing graph explainers (e.g., GNNExplainer) typically operate at the node/edge level, missing the holistic "pathway" structure and often failing to preserve long signaling chains.

Goal: To infer targeted, data-specific subgraphs from a static bio-network that capture the most biologically meaningful interactions driving a specific classification (e.g., disease type) based on experimental molecular data.

---

2. Methodology: The EXPATH Framework

The authors propose EXPATH, a two-stage deep learning framework consisting of a hybrid classifier (PATHMAMBA) and a pathway-level explainer (PATHEXPLAINER).

A. Data Encoding

• Input: A curated bio-network $G=(V, E)$ and condition-specific experimental data (e.g., amino acid sequences).
• Feature Extraction: The framework integrates biological foundation models. Specifically, it uses ESM-2 (a large protein language model) to encode amino acid sequences into node feature vectors, capturing complex structural and functional information.

B. PATHMAMBA: Pathway Representation Learning

To address the need for capturing both local interactions and global pathway dependencies, PATHMAMBA combines:

1. Local Aggregation (GIN): Uses a Graph Isomorphism Network (GIN) to aggregate information from immediate neighbors, ensuring high expressiveness for local structures.
2. Global Pathway Modeling (Mamba):
   • Random Pathway Sampling: Instead of processing the whole graph at once, the model samples random pathways (sequences of nodes) starting from each node.
   • State-Space Modeling (Mamba): Applies the Mamba architecture (a selective state-space model) to these sampled pathways. This allows the model to capture long-range dependencies and sequential logic within a pathway, mimicking biological signal propagation.
   • Integration: The outputs of the local GIN and the global Mamba are concatenated and processed through an MLP to generate the final graph representation for classification.

C. PATHEXPLAINER: Targeted Pathway Inference

To identify the specific subgraph responsible for the prediction:

• Pathway-Level Masking: Unlike standard explainers that mask individual edges, PATHEXPLAINER learns pathway masks. It treats entire pathways (connected sequences of nodes/edges) as units.
• Optimization Objective: It optimizes a mutual information objective to find the minimal subgraph $\hat{G}$ that preserves the prediction $y$.
  • Loss Function: Combines cross-entropy (prediction consistency) with an $L_1$ regularization term to enforce sparsity (selecting only the most critical pathways).
• Output: A subgraph representing the "targeted pathway" specific to the experimental data.

---

3. Key Contributions

1. Explicit Subgraph Inference Formulation: The paper reframes bio-network inference as a graph learning and explanation task, explicitly identifying subgraphs that contribute most to classification rather than just reconstructing the whole graph.
2. Novel Hybrid Architecture (PATHMAMBA): Introduces a model combining GNNs with State-Space Models (Mamba) and random pathway sampling. This theoretically overcomes the expressiveness limits of standard 1-WL GNNs, allowing the capture of higher-order structural patterns and long-range dependencies.
3. Pathway-Level Explainability (PATHEXPLAINER): Proposes a novel explainer that learns masks at the pathway level, ensuring the inferred subgraphs maintain coherent signaling chains rather than fragmented edges.
4. ML-Oriented Biological Evaluation: Introduces a new evaluation workflow that quantitatively assesses biological relevance using metrics like Fidelity+ (necessity), Fidelity- (sufficiency), and Enrichment Contribution Score (ECS), moving beyond qualitative expert review.

---

4. Experimental Results

The framework was evaluated on 301 human bio-networks from KEGG, categorized into four classes: Human Diseases, Metabolism, Molecular & Cellular Processes, and Organismal Systems.

A. Classification Performance (PATHMAMBA)

• Accuracy: PATHMAMBA achieved the highest accuracy (0.744) across all categories, outperforming baselines like GCN, GAT, GPS, and Graph-Mamba.
• Ablation: Removing the ESM-2 encoder dropped accuracy significantly (to 0.44), proving the necessity of foundation model embeddings for capturing biological nuance.
• Efficiency: Training was 30% faster than GPS, and inference was 27% faster than Graph-Mamba.

B. Pathway Inference Quality (PATHEXPLAINER)

• Fidelity:
  • Fidelity+ (Necessity): Achieved 4.5× higher scores than baseline explainers, meaning the extracted subgraphs are strictly necessary for the prediction.
  • Fidelity- (Sufficiency): Achieved 14× lower scores, meaning the subgraphs alone are sufficient to maintain prediction accuracy without the rest of the network.
• Path Preservation: The inferred subgraphs preserved signaling chains up to 4× longer (Max Path Length: 16 vs. ~4-6 in baselines) and had larger diameters, indicating better retention of long-range biological interactions.

C. Biological Meaningfulness

• Gene Ontology (GO) Analysis: The extracted subgraphs showed significantly higher Number of Enriched Biological Functions (#EBF) and Enrichment Contribution Scores (ECS) compared to statistical baselines (e.g., Random Sampling, Minimum Dominating Set).
• Case Study (TCR Signaling): In a T-cell receptor signaling case study, EXPATH correctly identified coherent, biologically validated signaling axes (PI3K-AKT and NF-κB), whereas baseline methods produced fragmented, scattered nodes that lacked mechanistic continuity.

---

5. Significance and Impact

• Bridging the Gap: EXPATH successfully bridges the gap between static biological knowledge bases and dynamic experimental data, allowing researchers to "zoom in" on relevant pathways for specific conditions.
• Interpretability: By preserving long signaling chains and providing pathway-level explanations, the model offers insights that are directly actionable for biologists, unlike "black-box" GNNs or fragmented edge-level explainers.
• Scalability: The integration of foundation models (ESM-2) and efficient state-space modeling (Mamba) demonstrates a scalable approach for analyzing complex biological systems without sacrificing computational efficiency.
• Future Applications: The framework is designed to be extensible to other bio-network types, promising broader applications in systems biology, drug discovery, and precision medicine.