Embedding Biological Knowledge into AI: A Path to Robust and Interpretable Precision Medicine
Loading...
Author
Meirer, Jonas
Poussin, Carine
DOI
Abstract
The increasing complexity of biomedical data has amplified the limitations of purely data-driven “black-box” machine learning models in clinical and translational research. This talk presents biologically informed neural networks (BINNs) as a principled solution to reconcile predictive performance with interpretability. By embedding curated biological knowledge—such as pathway hierarchies and molecular interaction networks—directly into AI model architectures, BINNs enable mechanistic insight across multiple biological scales. A DeepBINN case study on acute kidney injury classification demonstrates how prior knowledge improves robustness, model stability, and biological plausibility while achieving competitive predictive accuracy. The presentation further discusses methodological extensions, including custom ontologies and multi-modal omics integration, positioning BINNs as a key enabling technology for trustworthy AI in precision medicine. Overall, the work highlights how embedding biological knowledge into AI models is essential for regulatory acceptance, clinical translation, and meaningful decision support.
Publication Reference
Impact 2025, Innovation LSZ Event, Switzerland
Year
2025