[Paper Review] Relation-weighted Link Prediction for Disease Gene Identification

Srivamshi Pittala, William C. Koehler+5 more|arXiv (Cornell University)|Nov 10, 2020

Bioinformatics and Genomic Networks20 references4 citations

TL;DR

This paper proposes relation-weighted link prediction, a novel graph representation learning method that improves disease gene identification by assigning learnable weights to different relation types in a biomedical knowledge graph. By optimizing these weights to prioritize gene-disease relationships, the method achieves a 24.1% relative improvement over the closest SOTA competitor and outperforms Open Targets in precision for predicting Parkinson’s disease drug targets in clinical trials.

ABSTRACT

Identification of disease genes, which are a set of genes associated with a disease, plays an important role in understanding and curing diseases. In this paper, we present a biomedical knowledge graph designed specifically for this problem, propose a novel machine learning method that identifies disease genes on such graphs by leveraging recent advances in network biology and graph representation learning, study the effects of various relation types on prediction performance, and empirically demonstrate that our algorithms outperform its closest state-of-the-art competitor in disease gene identification by 24.1%. We also show that we achieve higher precision than Open Targets, the leading initiative for target identification, with respect to predicting drug targets in clinical trials for Parkinson's disease.

Motivation & Objective

• To address the challenge of identifying disease-associated genes, which are critical for understanding and treating complex diseases like Parkinson’s.
• To improve link prediction in heterogeneous biomedical knowledge graphs by accounting for imbalances in relation type frequencies.
• To develop a method that enhances prediction performance for gene-disease relationships, a key but underrepresented relation type in existing models.
• To empirically evaluate the impact of different relation types—such as protein-protein interactions, disease-ontology mappings, and gene-chemical associations—on prediction accuracy.
• To benchmark the proposed method against state-of-the-art approaches and Open Targets, the leading initiative in target identification, particularly for clinical trial relevance.

Proposed method

• Constructs a biomedical knowledge graph integrating multiple data sources, including gene-disease associations (DoidGeNET), protein-protein interactions (STRING), disease-gene associations from UniProt (DG_UC), disease ontology (DO), and drug-target relations (RT).
• Applies RotatE, a graph representation learning model that embeds entities and relations in complex vector space, modeling edges as rotations between head and tail entities.
• Introduces relation-weighted link prediction by modifying RotatE’s loss function to scale the contribution of each relation type by a learnable weight $ w_r $, optimizing for gene-disease link prediction performance.
• Uses negative sampling and margin-based loss to train the model, with the modified loss function $ L' = -\log\sigma(\gamma - w_r \cdot d_r(\mathbf{h}, \mathbf{t})) - \sum_{i=1}^{n} p(h_i', r, t_i') \log\sigma(d_r(\mathbf{h}_i', \mathbf{t}_i') - \gamma) $ to prioritize underrepresented but critical relation types.
• Trains and evaluates the model on a multi-layered graph, progressively adding relation types to assess their contribution to performance.
• Performs hyperparameter tuning and validation using standard metrics such as hit@k, mean rank, and mean percentile to ensure robust evaluation.

Experimental results

Research questions

• RQ1 How do different relation types in a biomedical knowledge graph contribute to the performance of disease gene prediction?
• RQ2 Can learning relation-specific weights improve link prediction performance for underrepresented but biologically critical relations like gene-disease?
• RQ3 Does relation-weighted link prediction outperform existing state-of-the-art methods in identifying disease genes?
• RQ4 How does the proposed method compare to Open Targets in predicting clinically relevant drug targets for Parkinson’s disease?
• RQ5 What is the impact of progressively enriching the knowledge graph with additional relation layers on prediction performance?

Key findings

• The addition of relation layers—STRING, DG_UC, DO, and RT—progressively improves performance, with the full graph achieving a hit@30 of 0.375, a mean rank of 1186.81, and a mean percentile of 93.32.
• Relation-weighted RotatE outperforms the original RotatE, improving hit@30 from 0.368 to 0.375 and reducing mean rank by 12.7%, demonstrating the effectiveness of relation-specific weighting.
• The proposed method achieves a 24.1% relative improvement in hit@100 over DIAMOnD, the closest SOTA competitor, with a hit@100 of 0.535 compared to DIAMOnD’s 0.431.
• In a direct comparison with Open Targets for predicting Parkinson’s disease drug targets in clinical trials, the model achieves higher precision across the top 50 predictions, with more cumulative hits and superior precision-recall trade-offs.
• The model achieves a 76.2% relative reduction in mean rank and a 98.4% relative increase in hit@30 when comparing the full graph to a minimal disease-gene-only graph, confirming the value of graph enrichment.
• The study confirms that relation-specific weighting significantly mitigates bias toward high-frequency relation types, improving performance on low-frequency but critical gene-disease links.