Researchers from Helmholtz Munich and the Mila Institute in Montréal have developed a new AI method called BioPathNet. It searches large biological data networks specifically for hidden relationships, from gene functions to disease mechanisms and therapeutic approaches. The method was published in the journal Nature Biomedical Engineering.

Biomedical knowledge graphs link genes, proteins, diseases, drugs and processes with each other. However, they are incomplete, and many suspected connections are missing. BioPathNet analyzes not only individual data points, but entire chains of relationships, for example from the gene to signaling pathways to diseases and drugs. From thousands of patterns, the model learns which new compounds are biologically plausible. The predictions are comprehensible, as the underlying paths in the network become visible for each recommendation.

In tests, BioPathNet predicted gene functions, disease relationships and therapeutic approaches, for example in leukemia, stomach cancer and Alzheimer’s disease. It rediscovered known therapies and proposed substances that are being tested in clinical trials. The method is not a black box model, but a tool for testable hypotheses. Any proposed compound must be tested experimentally or clinically. The quality depends on the input data.

In the long term, BioPathNet is intended to become a building block for comprehensive basic models of biomedical graphs that can be adapted for various tasks, from drug discovery to deciphering disease mechanisms. It is based on the graph neural network framework NBFNet and combines predictive power with interpretability. As an open-source solution, it is available to researchers worldwide.

The idea came about during a stay of a doctoral student at the Mila Institute. An originally planned project on air pollution failed due to a lack of data, so the team chose biomedical knowledge graphs as an alternative. The interdisciplinary collaboration of computational biology, mathematics, biophysics and computer science across locations was crucial.

BioPathNet helps to make better use of scattered knowledge in data networks and to generate new ideas for experiments and therapies. It addresses the complexity of biomedical data and supports hypothesis formation in research. By focusing on paths instead of isolated points, it improves the plausibility of predictions. In oncology and neurology, it could provide new approaches for personalized medicine.

The method reduces the effort required for manual analysis of large graphs and makes hidden patterns accessible. It complements existing AI approaches with increased transparency, which increases acceptance in medical research. Future developments may include the integration of other data sources to increase accuracy. The tool helps accelerate biomedical discoveries and could lead to new therapies in the long term. The cooperation between Helmholtz Munich and Mila shows the potential of international partnerships in AI research.

Original Paper:

Enhancing link prediction in biomedical knowledge graphs with BioPathNet | Nature Biomedical Engineering

Editor: X-Press Journalistenbüro GbR

Gender Notice. The personal designations used in this text always refer equally to female, male and diverse persons. Double/triple naming and gendered designations are used for better readability. ected.