Interpretable Deep Learning for Biomarker Discovery

NAISS 2023/22-417


NAISS Small Compute

Principal Investigator:

Sören Richard Stahlschmidt


Högskolan i Skövde

Start Date:


End Date:


Primary Classification:

10203: Bioinformatics (Computational Biology) (applications to be 10610)



Deep neural networks (DNNs) are powerful function approximators and over the last years have enabled rapid progress in machine learning. Predicting biomedical endpoints has been no exception to this trend. However, in order to discover novel biomedical knowledge predictions by machine learning models must be explainable. Due to their complexity and distributed representations DNNs are difficult to interpret. In this project we want to develop novel explainability methods and evaluate existing ones for their potential to generate novel biological knowledge. In particular, we aim to investigate these methods with the cases of single cell RNA-seq of cardiac hypertrophy in vitro models and breast cancer in vivo samples as well as bulk DNA methylation and gene expression data of cancer patients. Additionally, we aim to investigate the interaction between discovered genetic biomarkers and morphological characteristics of cells observable on whole slide images of tumor biopsies. The results of this project will contribute to making DNNs incraesingly useful for biomedical knowledge generation.