The increasing use of neural networks in biological and medical data analysis raises fundamental challenges related to interpretability, generalization, and biological plausibility. Classical purely data-driven deep learning models often fail to align with existing biological knowledge, and their predictions can be difficult to explain in a biologically meaningful way.

Biologically Informed Neural Networks (BINNs) aim to explicitly incorporate prior biological knowledge into the learning process, including gene–gene interactions, signaling pathways, protein–protein interaction networks, and hierarchical biological ontologies. By embedding such structured information into the model architecture or training objective, BINNs can achieve improved interpretability, enhanced robustness, and better performance in low-sample-size regimes.

The goal of this project is to design, implement, and apply a biologically informed neural network model using real-world, publicly available biological datasets. The student is expected to construct a model whose architecture or regularization reflects biological prior knowledge (e.g., pathways, networks, or ontological hierarchies), and to apply it to a relevant prediction task such as disease outcome, phenotype prediction, or drug response.

A key component of the project is the application of model interpretability and explainability methods (e.g., feature attribution, pathway-level aggregation, gradient-based or perturbation-based approaches), and the systematic comparison of results with standard, non-biologically informed neural network models. The analysis should critically assess how biological priors affect predictive performance, stability, and interpretability.

Nature of the tasks:

• Processing and integration of public biological datasets

• Formal representation of biological prior knowledge (networks, hierarchies, masks, regularization terms)

• Design and implementation of biologically informed neural network architectures

• Model training, testing, and quantitative evaluation

• Application of explainability methods and comparative analysis

• Critical interpretation of results from a biological perspective

Framework:

• PyTorch

• (optionally) PyTorch Geometric, NumPy, pandas