Link prediction in biological knowledge graphs. Credit: Nature Biomedical Engineering (2026). DOI: 10.1038/s41551-025-01598-z. https://www.nature.com/articles/s41551-025-01598-z

A new artificial intelligence (AI) method called BioPathNet helps researchers systematically search large biological data networks for hidden connections—from gene functions and disease mechanisms to potential therapeutic approaches. BioPathNet was developed by teams at Helmholtz Munich and Mila—Quebec Artificial Intelligence Institute in Montreal, Canada. The researchers have now published the method in the journal Nature Biomedical Engineering.

Interpretable AI connects genes, diseases, and drugs

Biomedical "knowledge graphs" are structured maps that link genes, proteins, diseases, drugs, and biological processes, capturing their relationships to help both humans and AI understand biology and medicine in a systematic way. However, these networks remain incomplete: Many relationships that researchers expect to exist have not yet been documented.

BioPathNet is an AI method for analyzing such biomedical knowledge graphs. Unlike many existing approaches, it does not just look at individual data points but at entire chains of relationships—for example, from a gene, via a signaling pathway, to a disease and on to a potential drug.

"From thousands of these patterns, the model learns which new, biologically plausible connections are likely," says Emy Yue Hu, first author of the study and a Ph.D. student at the Institute of Computational Biology (ICB) at Helmholtz Munich. "On this basis, BioPathNet proposes hypotheses that we can then test in the lab or in clinical studies."

A key advantage of BioPathNet is that its predictions are interpretable: For each suggestion, the model can trace which paths in the knowledge graph led to that prediction.

Turning data points into testable hypotheses

In extensive tests, the team led by Dr. Annalisa Marsico, group leader at ICB and principal investigator of the project, applied BioPathNet to a wide range of tasks—predicting gene functions, uncovering relationships between diseases, identifying potential targets for cancer therapy and suggesting new indications for established drugs.

For complex diseases such as leukemia, gastric cancer and Alzheimer’s disease, BioPathNet not only rediscovered known therapies but also highlighted compounds that are already being tested in clinical trials. "It was crucial for us not to build yet another black-box model," stresses Marsico. "For every prediction we can inspect the most important paths in the graph and discuss with domain experts whether they make biological sense."

BioPathNet is therefore not an automatic recommendation engine for therapies, but a hypothesis-generating tool. The quality of its suggestions depends on the underlying data, and every predicted connection must be validated experimentally or clinically. In the long term, the researchers see the method as a building block towards foundation models for biomedical knowledge graphs that can be fine-tuned for many different tasks—from drug repurposing and the dissection of disease mechanisms to applications beyond medicine.

"Our goal is not a miracle cure powered by AI," says Marsico. "We want a tool that helps us make better use of existing biomedical data networks and come up with good new ideas for experiments and therapies."

Bridging disciplines to build BioPathNet

The idea for BioPathNet emerged during a research stay by Hu at Mila. She had originally planned to work there on an air pollution project, but the available data turned out not to be suitable. Together with her supervisors, she looked for a new, data-driven research question—and arrived at large biomedical knowledge graphs, which already contain a great deal of scattered information on genes, diseases and therapies.

The team’s expertise in machine learning at Mila was crucial for developing the algorithms underlying BioPathNet, which is built on the NBFNet graph neural network framework, ensuring both predictive power and interpretability.

"BioPathNet was only possible because of the interdisciplinarity of the team," says Hu. "Experts in computational biology, mathematics, biophysics and computer science worked closely together across the sites in Montreal and Munich."

As an open-source tool, BioPathNet is now available to researchers worldwide to explore biomedical mechanisms.

Publication details

Emy Yue Hu, et al, Enhancing link prediction in biomedical knowledge graphs with BioPathNet, Nature Biomedical Engineering (2026). DOI: 10.1038/s41551-025-01598-z. www.nature.com/articles/s41551-025-01598-z

Journal information: Nature Biomedical Engineering

Citation: New AI uncovers hidden patterns in biomedical knowledge graphs (2026, January 20) retrieved 20 January 2026 from https://medicalxpress.com/news/2026-01-ai-uncovers-hidden-patterns-biomedical.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.