[Paper Review] KGML-xDTD: A Knowledge Graph-based Machine Learning Framework for Drug Treatment Prediction and Mechanism Description

Chunyu Ma, Zhihan Zhou+2 more|arXiv (Cornell University)|Nov 30, 2022

Computational Drug Discovery Methods4 citations

TL;DR

KGML-xDTD is a knowledge graph-based machine learning framework that predicts drug-disease treatment relationships and provides biologically interpretable, path-based mechanisms of action (MOAs) using graph-based reinforcement learning. It achieves state-of-the-art performance in drug repurposing prediction and MOA path recapitulation on the large-scale RTX-KG2c biomedical knowledge graph.

ABSTRACT

Background: Computational drug repurposing is a cost- and time-efficient approach that aims to identify new therapeutic targets or diseases (indications) of existing drugs/compounds. It is especially critical for emerging and/or orphan diseases due to its cheaper investment and shorter research cycle compared with traditional wet-lab drug discovery approaches. However, the underlying mechanisms of action (MOAs) between repurposed drugs and their target diseases remain largely unknown, which is still a main obstacle for computational drug repurposing methods to be widely adopted in clinical settings. Results: In this work, we propose KGML-xDTD: a Knowledge Graph-based Machine Learning framework for explainably predicting Drugs Treating Diseases. It is a two-module framework that not only predicts the treatment probabilities between drugs/compounds and diseases but also biologically explains them via knowledge graph (KG) path-based, testable mechanisms of action (MOAs). We leverage knowledge-and-publication based information to extract biologically meaningful "demonstration paths" as the intermediate guidance in the Graph-based Reinforcement Learning (GRL) path-finding process. Comprehensive experiments and case study analyses show that the proposed framework can achieve state-of-the-art performance in both predictions of drug repurposing and recapitulation of human-curated drug MOA paths. Conclusions: KGML-xDTD is the first model framework that can offer KG-path explanations for drug repurposing predictions by leveraging the combination of prediction outcomes and existing biological knowledge and publications. We believe it can effectively reduce "black-box" concerns and increase prediction confidence for drug repurposing based on predicted path-based explanations, and further accelerate the process of drug discovery for emerging diseases.

Motivation & Objective

• Address the critical need for explainable, reliable drug repurposing predictions in emerging and orphan diseases.
• Overcome the 'black-box' limitation of existing machine learning models in drug repurposing by providing testable, biologically meaningful mechanisms of action (MOAs).
• Enable clinical adoption of computational drug repurposing by integrating existing biomedical knowledge and publications into interpretable prediction paths.
• Scale drug repurposing prediction to massive, complex biomedical knowledge graphs (e.g., RTX-KG2c) with millions of nodes and edges.
• Improve prediction confidence by linking model outputs to known biological pathways and literature-curated MOA evidence.

Proposed method

• Construct a two-module framework: (1) a drug-disease treatment prediction module using graph-based reinforcement learning (GRL), and (2) a mechanism explanation module using knowledge graph (KG) path extraction.
• Leverage knowledge-and-publication-based information to extract biologically meaningful 'demonstration paths' as intermediate guidance in the GRL process.
• Use a large-scale, multi-source biomedical knowledge graph (RTX-KG2c) integrating data from databases, literature, and EHRs to represent drug, disease, gene, and protein relationships.
• Train the GRL agent to discover 3-hop paths in the KG that connect drugs to diseases, with rewards shaped by biological plausibility and literature support.
• Integrate external biological databases (e.g., DrugMechDB) and case studies to validate predicted MOA paths against known mechanisms.
• Apply knowledge graph completion techniques (e.g., TransE, RotatE) as baselines for comparison in prediction performance.

Experimental results

Research questions

• RQ1 Can a knowledge graph-based machine learning framework predict drug-disease treatment relationships with higher accuracy and lower false positives than existing state-of-the-art models on large-scale biomedical knowledge graphs?
• RQ2 Can the framework generate biologically interpretable, testable, and literature-supported mechanisms of action (MOAs) for predicted drug-disease pairs?
• RQ3 To what extent can the predicted KG paths recapitulate human-curated, experimentally validated drug MOA pathways?
• RQ4 How does the integration of knowledge-and-publication-based demonstration paths improve the interpretability and reliability of drug repurposing predictions?
• RQ5 Can the framework be effectively scaled to massive, complex biomedical knowledge graphs (e.g., RTX-KG2c) while maintaining high performance and explainability?

Key findings

• KGML-xDTD achieved state-of-the-art performance in drug repurposing prediction on the large-scale RTX-KG2c biomedical knowledge graph, outperforming existing models in accuracy and reducing false positives.
• The framework successfully recapitulated known biological mechanisms: for Huntington’s disease, it predicted that Isradipine acts via calcium voltage-gated channels (CACNA1S, CACNA1D, etc.), aligning with literature reports on its potential therapeutic role.
• For Huntington’s disease, the model predicted that Amifampridine may regulate potassium voltage-gated channels, a mechanism supported by prior studies (Noh et al., 2019).
• Top predicted drugs for Huntington’s disease (e.g., Pimozide, Olanzapine, Riluzole) had high prediction probabilities (0.939, 0.938, 0.935) and were supported by clinical literature.
• The predicted 3-hop MOA paths for non-chemotherapeutic drugs were biologically plausible and integrated into subgraphs that reflect known regulatory networks.
• Validation using DrugMechDB confirmed that the predicted MOA paths were consistent with known drug mechanisms, demonstrating the model’s ability to generate testable, biologically meaningful explanations.