🧬 biology

PPI-Net connects molecular protein interactions to functional processes in disease

The paper introduces PPI-Net, a hierarchical graph neural network that integrates protein-protein interaction networks with pathway-level hierarchies to accurately predict cancer states and provide mechanistic insights into disease-driving biological processes.

Original authors: Kyle Higgins, Guadalupe Gonzalez, Dennis Veselkov, Ivan Laponogov, Kirill Veselkov

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kyle Higgins, Guadalupe Gonzalez, Dennis Veselkov, Ivan Laponogov, Kirill Veselkov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to understand why a city (a human body) has fallen into chaos (cancer). You have a massive list of individual problems: a broken traffic light here, a pothole there, a missing street sign over there.

Most computer programs trying to diagnose cancer look at this list of broken parts one by one. They might say, "Oh, this traffic light is broken, so it's a problem!" But they often miss the bigger picture: how that broken light causes a traffic jam, which then causes a delivery truck to get stuck, which then stops food from reaching a neighborhood. They see the pieces, but not the story.

The paper you shared introduces a new tool called PPI-Net. Think of PPI-Net as a smart city planner that doesn't just look at the broken parts, but understands the entire map of how the city is connected.

Here is how it works, using simple analogies:

1. The Two Maps (The Ingredients)

To understand the city, PPI-Net uses two specific maps:

The "Who Knows Whom" Map (PPI Network): This shows which proteins (the city workers) physically shake hands or talk to each other. If Worker A is sick, PPI-Net knows exactly which neighbors (Workers B and C) will be affected immediately.
The "Department Hierarchy" Map (Reactome Pathways): This is like the city's organizational chart. It groups workers into teams (e.g., "The Fire Department," "The Waste Management Team," "The Power Grid"). It shows how a small team fits into a larger department, which fits into the whole city government.

2. The Conveyor Belt (How It Works)

Most AI models try to guess the disease by looking at the "broken parts" directly. PPI-Net is different. It uses a "conveyor belt" system:

Step 1: The Local Neighborhood: First, it looks at the patient's specific "broken parts" (their genetic data) and places them on the "Who Knows Whom" map. It sees how the trouble spreads locally among neighbors.
Step 2: The Assembly Line: Then, it puts that local trouble onto a conveyor belt that moves up the "Department Hierarchy."
- It takes the trouble from individual workers and asks, "What is this doing to the Fire Department?"
- Then it asks, "What is the Fire Department's trouble doing to the City Safety level?"
- Finally, it asks, "How does this affect the Whole City?"

By moving the information up this ladder, the model can see if a small problem in a single protein is actually causing a massive failure in a whole biological system (like stress response or metabolism).

3. The "Team Captain" Check (Why It's Better)

The researchers tested this on breast cancer and other types. They found that PPI-Net is much better at guessing if a patient has cancer than models that just look at the broken parts alone.

The "PPI-Only" Model: This is like a detective who only looks at the broken traffic lights. It's okay, but it misses the big traffic jam.
The "Top-Down Only" Model: This is like a manager who only looks at the final report from the Mayor without checking the teams. It's also okay, but it misses the details.
PPI-Net: This model checks the broken lights, sees how they jam the traffic, checks the Fire Department's response, and then looks at the Mayor's report.

The Result: By using this "conveyor belt" approach, PPI-Net improved its accuracy by about 6.7% compared to just looking at the broken parts, and by 12.3% compared to just looking at the final report. It was so good that in many cases, it correctly identified cancer over 90% of the time.

4. The "Flashlight" (What It Tells Us)

One of the coolest things about PPI-Net is that it doesn't just give a "Yes/No" answer; it shines a flashlight on why it made that decision.

When they looked at the results for breast cancer, the model highlighted specific "teams" that were acting up:

The Stress Response Team: It found that the cell's "stress alarms" (involving proteins like TP53 and HSP90) were screaming. This is a known way cancer survives.
The "Ion Channel" Team: Surprisingly, the model also highlighted a group of proteins that act like electrical switches (ion channels) in the cell. While scientists knew these were important, the model showed they were working together in a coordinated way to help the cancer grow. It's like realizing the city's power grid is being hijacked to keep the chaos going.

The Bottom Line

The paper claims that PPI-Net is a new way to use AI to study cancer. Instead of treating a patient's biology as a random list of broken genes, it treats it like a connected city with a clear hierarchy.

By connecting the small, local problems (molecular interactions) to the big, organized teams (pathways), the model can:

Predict cancer more accurately.
Explain its answer by showing exactly which biological "teams" are malfunctioning.

The authors say this helps us understand how cancer works, not just that it exists, by revealing how small molecular changes ripple up to cause big system failures.

Technical Summary of PPI-Net: Connecting Molecular Protein Interactions to Functional Processes in Disease

Problem Statement
Understanding how molecular alterations propagate across biological systems to drive disease remains a central challenge in cancer biology. While high-throughput profiling (e.g., RNA-seq, DNA methylation) enables comprehensive characterization of tumor states, most existing computational models treat these data as independent features. This approach neglects the structured biological relationships that govern cellular function, resulting in models that may achieve strong predictive performance but lack interpretability and mechanistic insight. Furthermore, existing network-based approaches typically apply Protein-Protein Interaction (PPI) networks or functional pathway annotations in isolation, failing to capture the full cascade from molecular perturbation to system-level dysfunction.

Methodology
The authors introduce PPI-Net, a hierarchical graph neural network (GNN) framework designed to bridge molecular interaction networks with pathway-level representations. The architecture operates in two sequential stages:

Patient-Specific PPI Graph Construction:
- Patient molecular data (RNA-seq and DNA methylation) from The Cancer Genome Atlas (TCGA) are mapped onto a fixed STRING PPI network.
- Nodes represent proteins (mapped to HGNC gene symbols), and edges represent experimentally supported physical interactions (top 100,000 confidence scores).
- A Graph Convolutional Network (GCN) processes these patient-specific graphs to propagate information across the interaction network, generating latent node embeddings that capture local molecular context.
Hierarchical Pathway Propagation (Reactome):
- Gene-level embeddings are mapped onto a curated Reactome pathway hierarchy, structured as a directed acyclic graph with five pathway layers and a terminal gene layer.
- A hierarchical Graph Attention Network (GAT) propagates signals in a bottom-up manner, from the gene layer through successive pathway levels to a synthetic root node.
- Bipartite Graph Attention: At each layer transition, attention-based message passing learns weighted contributions of lower-level nodes to higher-level pathway representations.
- Multi-Level Supervision: The model generates a scalar prediction logit at each hierarchical level. These logits are aggregated using a geometrically decaying weighting scheme (controlled by parameter $\alpha$ ) to produce the final classification output. This allows the model to integrate information across multiple biological scales while prioritizing higher-level abstractions.
Explainability Framework:
- The model employs a multi-level attribution framework combining gradient-based importance with attention-driven information flow.
- PPI Level: Node importance is computed via gradients; edge importance is derived from node scores to identify high-impact interaction subgraphs.
- Pathway Level: A "directed flow" metric is defined as the product of attention coefficients and destination node importance, quantifying signal transmission across the hierarchy.

Key Contributions

Architectural Innovation: PPI-Net introduces a "conveyor-belt" design that treats pathway hierarchy levels as node types and biological parent-child relationships as typed edges. This converts curated pathway hierarchies into graph-native computational objects, enabling attention-based message passing across biological scales.
Multi-Scale Integration: Unlike prior models that focus on either molecular graphs or pathway abstractions, PPI-Net explicitly links the two, capturing local molecular context before aggregating signals into higher-order biological programs.
Interpretability: The framework provides interpretable, multi-level attributions that connect model predictions to biologically meaningful mechanisms, recovering known oncogenic modules while highlighting less characterized signals.

Results
The model was evaluated across ten cancer types (including BRCA, LUAD, LUSC, LIHC) using TCGA datasets.

Predictive Performance: PPI-Net achieved robust balanced accuracy (BA) exceeding 90% in multiple cohorts (e.g., LUAD: 93.4%, LUSC: 96.2%, BRCA: 91.4%).
Ablation Studies (TCGA-BRCA):
- Reactome Hierarchy: Removing the Reactome hierarchy (PPI-only model) reduced BA from 0.93 to 0.87, a relative improvement of 6.7% attributed to pathway-level organization.
- Multi-Level Supervision: Removing intermediate hierarchical supervision (Top-layer-only loss) reduced BA to 0.83 with higher variance. The full multi-level strategy provided a 12.3% relative improvement, indicating that supervision at intermediate scales stabilizes training and improves generalization.
- Multi-Omics: Integrating RNA-seq with DNA methylation data improved validation BA to 98.0% and test BA to 88.3% in breast cancer.
Biological Insights:
- The model successfully recovered canonical oncogenic modules, including the TP53–MDM2–AKT1 regulatory axis and HSP90-mediated proteostasis.
- It identified strong convergence onto ion signaling (potassium and calcium channels, e.g., KCNK, KCNJ families) and cellular responses to stimuli, linking these to neuronal system pathways and stress responses.
- Multi-omics integration sharpened representations toward disease-relevant regulatory programs, contextualizing transcriptional signals within epigenetic frameworks.

Significance
The paper claims that integrating interaction networks with pathway hierarchies enables accurate prediction while providing mechanistic insight into cancer biology. By structuring the model around biologically meaningful relationships, PPI-Net reduces reliance on purely data-driven feature learning, constraining representations to reflect known biological organization. This approach offers a pathway toward more transparent and biologically grounded machine learning models in cancer research, capable of tracing predictions from proteins to genes and pathways to generate testable hypotheses. The authors note that while the model relies on curated databases and fixed structures, it demonstrates the value of embedding pathway hierarchies into predictive models to capture multi-scale biological organization.