Uncertainty-aware graph representation learning with positive-unlabeled classification for biomarker discovery in peripheral artery disease

This paper presents an uncertainty-aware graph representation learning framework that integrates positive-unlabeled classification and ensemble methods to prioritize novel and well-calibrated biomarkers for peripheral artery disease, demonstrating superior predictive performance and biological relevance compared to existing baselines.

The Gist

Imagine Peripheral Artery Disease (PAD) as a massive, tangled city of roads (our blood vessels) where some streets are blocked, but we don't have a complete map of why they are blocked. Scientists know a few key culprits (the "positive" proteins), but for most of the city, the traffic reports are missing or incomplete. This makes it very hard to find new clues to fix the problem.

Most computer programs trying to solve this are like overconfident tour guides. They point at a random building and say, "This is definitely the problem!" without admitting they might be guessing. They don't know when they are unsure, and they often miss new, strange buildings that don't look like the ones they've seen before.

The New Approach: A Cautious Detective with a Crystal Ball

The researchers in this paper built a smarter system, like a cautious detective who carries a "confidence meter" and a "novelty radar." Here is how they did it:

1. Mapping the City (Graph Learning): First, they used a special type of AI (a Graph Neural Network) to create a 3D map of how all the proteins in the body connect to each other. Think of this as drawing a subway map where the distance between stations represents how closely related different proteins are.
2. The "Yes, Maybe, No" Team (Ensemble Prediction): Instead of relying on one single detective, they hired a whole team of different experts (five different classifiers) and asked them to vote. They also taught these experts to say, "I'm not sure," when the data was fuzzy. This created a "confidence meter" that tells us how sure the system is about its answer.
3. The Two Buckets (Uncertainty & Novelty): The system sorted the potential clues into two piles:
   • The "Safe Bets": These are candidates that look very similar to the known troublemakers. The system is very confident about these.
   • The "New Discoveries": These are candidates that live in strange, unexplored parts of the map. The system flags these as "structurally novel" because they don't fit the usual patterns, suggesting they might be new types of culprits we haven't thought of yet.

What They Found

The team tested this system and found it was much better than the old methods. While the old "overconfident" guides were right about 82% of the time, this new team was right about 92% of the time.

• The Safe Bets: The proteins the system was most confident about clustered together with known PAD proteins. They were involved in familiar tasks like building the road walls (extracellular matrix) and managing blood clotting (coagulation).
• The New Discoveries: The "novel" candidates lived in different neighborhoods on the map. These were linked to different kinds of traffic control, like cell signaling and immune system responses (G protein-coupled receptors and NF-kappaB pathways).

The Bottom Line

By teaching the computer to admit when it's unsure and to look for things that are different from the norm, the researchers successfully identified 100 new potential biomarkers for PAD. They proved that mixing "confidence" with "curiosity" helps scientists find both the obvious suspects and the hidden ones, leading to a much clearer picture of the disease.

Technical Summary

Technical Summary: Uncertainty-aware Graph Representation Learning for PAD Biomarker Discovery

Problem Statement
Peripheral artery disease (PAD) presents a significant challenge for systematic biomarker discovery due to its heterogeneous molecular mechanisms and the incomplete functional annotation of relevant genes. While network-based learning approaches offer a powerful framework for disease gene prioritization, existing methods suffer from critical limitations: they typically generate overconfident predictions without explicitly modeling uncertainty, and they fail to account for structural novelty within the biological network. This lack of calibration and novelty detection hinders the reliable identification of both known associations and potentially novel therapeutic targets.

Methodology
To address these gaps, the authors propose an uncertainty-aware framework that integrates four core components: unsupervised graph representation learning, positive-unlabeled (PU) classification, ensemble prediction, and mechanistic explainability.

1. Graph Representation Learning: Node embeddings are generated using multiple unsupervised graph neural network (GNN) objectives. This multi-objective approach captures diverse structural properties of the biological network.
2. Ensemble PU Classification: These embeddings are fed into a heterogeneous set of five classifiers. The system employs an ensemble strategy to aggregate predictions, generating probability estimates and quantifying epistemic uncertainty.
3. Candidate Stratification: The framework jointly models predictive confidence and embedding-space novelty. This allows candidates to be stratified into two distinct categories:
   • High-confidence rediscoveries: Candidates with high predictive probability that cluster tightly with known PAD proteins.
   • Structurally novel hypotheses: Candidates that occupy distinct regions in the embedding space, representing potential new biological mechanisms, all under explicit uncertainty control.
4. Validation: The system utilizes five-fold cross-validation against established PU learning baselines and performs external validation using cardiovascular disease annotations.

Key Results
The proposed ensemble approach demonstrated robust performance across eight embedding objectives and five classifiers:

• Performance Metrics: The method achieved a consistent improvement over the best baseline PU learning method, yielding an Area Under the Curve (AUC) of 0.916 ± 0.019 compared to 0.821 ± 0.030 for the baseline.
• Candidate Prioritization: The model successfully prioritized 100 candidate PAD-associated proteins.
• Biological Enrichment:
  • Probability-heavy candidates (high confidence) were enriched for established vascular and hemostatic pathways, including extracellular matrix organization, integrin signaling, coagulation, and fibrinolysis.
  • Novelty-heavy candidates (structurally distinct) partitioned into coherent clusters enriched for upstream regulatory and signaling processes, specifically G protein-coupled receptor, ephrin receptor, kinase-driven, and NF-κB-associated pathways.
• External Validity: Top candidates showed significant enrichment for related cardiovascular disease annotations, performing 5.7 times above the background level.

Significance and Claims
The paper claims that integrating uncertainty quantification, novelty detection, and explainability enables a more calibrated and biologically grounded approach to biomarker prioritization. By moving beyond overconfident predictions, the framework not only recovers known vascular mechanisms but also systematically identifies structurally novel hypotheses with controlled uncertainty. The authors assert that this approach offers broad applicability not only to PAD but to other complex diseases characterized by heterogeneous mechanisms and incomplete annotations, providing a robust tool for distinguishing between established associations and genuine novel discoveries.