🧬 biology

XG-PUL: A Network-Based Framework for Disease Gene Prioritization Using Graph Representation Learning and Positive-Unlabeled Learning

The paper proposes XG-PUL, a robust machine learning framework that combines Node2Vec-based graph representation learning with a bagging-based positive-unlabeled classifier to effectively prioritize disease-associated genes by addressing the challenge of label uncertainty in biological networks.

Original authors: Mahdiyeh Ayouman, Zahra Narimani

Published 2026-08-12

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Mahdiyeh Ayouman, Zahra Narimani

Original paper licensed under CC BY 4.0 (https://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 the human body as a massive, bustling city. In this city, every gene is a citizen, and the proteins they make are the tools and messages they use to talk to one another. These conversations form a giant web of connections called a "protein-protein interaction network." Sometimes, a few citizens get sick, and their bad habits spread through the city, causing a disease. Scientists call these sick clusters "disease modules." The big challenge? We know the names of a few "bad apples" (confirmed disease genes), but we have no idea who the rest of the troublemakers are. The rest of the city's population is just a big, unlabeled crowd. If we try to guess who is sick by assuming everyone else is healthy, we might get it wrong because some of those "healthy" people are actually sick but haven't been caught yet. This is the tricky puzzle of finding disease genes: how do you find the hidden troublemakers when you can't be sure who the good guys are?

This is where a new tool called XG-PUL steps in, acting like a super-smart detective for the city's web. The researchers, Mahdiyeh Ayouman and Zahra Narimani, built a computer framework to solve this specific problem. Instead of guessing who is healthy, they used a clever strategy called "Positive-Unlabeled learning." Think of it like a game of "hot and cold." You know where the treasure is (the confirmed disease genes), and you have a map of the city (the protein network), but you don't know where the other treasures are hidden. XG-PUL first learns the "personality" of every citizen just by looking at who they hang out with and how central they are in the city, without needing to know if they are sick or healthy yet. Then, it uses a team of detectives (an ensemble of classifiers) to vote on who is most likely to be part of the sick cluster. By having many detectives look at slightly different groups of people and averaging their votes, the system becomes very good at spotting the hidden troublemakers, even if some of them are pretending to be healthy.

The team tested this detective on ten different complex diseases, including breast cancer, schizophrenia, and liver cirrhosis. They compared XG-PUL against other famous methods that have been used for years. The results were promising: XG-PUL was better at ranking the most likely disease genes at the top of the list. For example, when looking for genes related to breast cancer, the system successfully placed known, biologically important genes near the very top, proving it wasn't just guessing randomly. It also found genes involved in specific pathways, like those controlling how cells talk to each other or how the immune system fights back.

However, the paper is careful not to claim this is a magic bullet that solves everything. The authors note that their method relies on the current maps of the city, which are still incomplete. If the map is missing a whole neighborhood, the detective can't find the troublemakers hiding there. Also, the current version looks at the city as a whole, rather than checking specific neighborhoods (like just the brain or just the liver), which might miss some very specific details. But for now, XG-PUL suggests that by combining a smart way of reading the city's map with a team-voting system, we can find new clues about what causes complex diseases, helping scientists focus their real-world experiments on the most promising suspects.

Technical Summary of XG-PUL: A Network-Based Framework for Disease Gene Prioritization

Problem Statement
The identification of disease-associated genes is critical for drug target discovery and biomarker development, yet experimental validation remains costly and time-consuming. While computational methods based on biological networks offer an alternative, they face a fundamental challenge: the lack of reliable negative examples. In disease gene prioritization, only a limited set of genes is confirmed as disease-associated (positive class), while the vast majority of remaining genes are unlabeled rather than confirmed negatives. Treating these unlabeled genes as reliable negatives introduces systematic bias and label uncertainty, degrading the performance of standard supervised learning models. Existing approaches often struggle to handle this "positive-unlabeled" (PU) setting effectively, particularly when integrating complex network topologies.

Methodology
The authors propose XG-PUL, a machine learning framework designed to address disease gene prioritization under PU learning conditions using Protein-Protein Interaction (PPI) networks. The framework operates through a multi-stage pipeline:

Data Curation: The study utilizes a human PPI network derived from BioGRID (containing ~19,761 genes and ~682,195 interactions) and gene-disease associations from DisGeNET. Ten complex diseases were analyzed, including breast cancer, schizophrenia, and colorectal carcinoma.
Graph Representation Learning: To capture latent structural relationships without relying on class labels, the framework employs Node2Vec to generate low-dimensional node embeddings. The authors systematically evaluated four embedding configurations (varying dimensionality and walk parameters) and selected the HighDim configuration ( $d=256$ ) based on its stable performance across diverse diseases.
Feature Integration: The Node2Vec embeddings are combined with explicit graph-topological features, including degree centrality, betweenness centrality, PageRank, closeness centrality, and clustering coefficient. A Random Forest classifier is used to rank features by impurity-based importance scores, and the top 150 most informative features are retained to balance predictive performance and computational efficiency.
Bagging-Based PU Learning: To handle the absence of confirmed negative samples, XG-PUL employs a bagging-based PU classifier. In this strategy:
- Known disease genes (seed genes) serve as the positive training set.
- The unlabeled gene pool is repeatedly resampled via bootstrapping to create surrogate negative sets.
- Multiple classifiers are trained on these bootstrap samples.
- Final predictions are aggregated (e.g., via probability averaging) to reduce the bias introduced by treating undiscovered disease genes as negatives.
- The base learner used for these classifiers is XGBoost.
Gene Ranking: The ensemble model assigns disease-association scores to all candidate genes, generating a ranked list for each target disease.

Key Results
XG-PUL was evaluated against representative network-based and graph learning baselines, including DIAMOnD, Random Walk with Restart (RWR), MCL, and variants of XGDAG (GNNExplainer and GraphSVX).

Numerical Performance: Across ten complex diseases, XG-PUL consistently achieved superior performance in AUPR (Area Under the Precision-Recall Curve) and F1-score at K=500. For instance, in breast cancer, XG-PUL achieved an AUPR of 0.6522 and an F1@500 of 0.1768, outperforming the next best method (XGDAG-GraphSVX) which scored 0.5442 and 0.1152, respectively. It also maintained competitive or superior AUC-ROC scores.
Robustness: The framework demonstrated consistent performance across varying ranking thresholds (K), indicating its ability to prioritize relevant genes effectively across different candidate list sizes.
Biological Validation: Functional enrichment analysis using Enrichr revealed that top-ranked candidates were significantly enriched in disease-relevant pathways (e.g., neuronal signaling for schizophrenia, immune regulation for breast cancer). Literature review of the top 10 candidates for specific diseases identified known disease-associated genes (e.g., MAP2K6, USP15 for breast cancer; GRID2, ITGB2 for schizophrenia) appearing near the top of the prioritized lists, confirming the model's ability to recover biologically meaningful targets.

Key Contributions

Novel Framework: The introduction of XG-PUL, which uniquely integrates label-independent network embeddings (Node2Vec) with explicit topological features within a Positive-Unlabeled learning setting.
Handling Label Uncertainty: The application of a bagging-based PU strategy using XGBoost to mitigate the bias caused by the lack of reliable negative samples, a common bottleneck in disease gene discovery.
Feature Synergy: Demonstration that combining latent network embeddings with explicit topological descriptors yields better predictive performance than using embeddings alone.
Comprehensive Evaluation: A rigorous benchmark across ten diverse complex diseases, showing consistent outperformance over state-of-the-art network-based and graph learning methods.

Significance and Claims
The paper claims that XG-PUL provides a robust framework for disease gene prioritization in scenarios characterized by incomplete biological knowledge. By avoiding the assumption that unlabeled genes are negative, the method reduces systematic bias and improves the recovery of true disease-associated genes. The authors assert that their approach effectively identifies genes involved in disease-related pathways and molecular mechanisms beyond current annotations.

The study concludes that integrating network representation learning with PU learning offers a viable strategy for discovering novel disease-associated genes. However, the authors modestly acknowledge limitations, noting that the framework relies on static, context-independent PPI networks and curated databases, which may introduce bias toward well-studied genes and fail to capture tissue-specific or subtype-specific variations. They suggest that future work could incorporate multi-omics data and context-specific expression profiles to further refine disease module identification.

More like this