No abstract

Machine learning (ML) algorithms are used to build predictive models or classifiers for specific disease outcomes using transcriptomic data. However, some of these models show deteriorating performance when tested on unseen data which undermines their clinical utility. In this study, we show the importance of directly embedding prior biological knowledge into the classifier decision rules to build simple and interpretable gene signatures. We tested this in two important classification examples: a) progression in non-muscle invasive bladder cancer; and b) response to neoadjuvant chemotherapy (NACT) in triple-negative breast cancer (TNBC) using different ML algorithms. For each algorithm, we developed two sets of classifiers: agnostic, trained using either individual gene expression values or the corresponding pairwise ranks without biological consideration; and mechanistic, trained by restricting the search to a set of gene pairs capturing important biological relations. Both types were trained on the same training data and their performance was evaluated on unseen testing data using different methodologies and multiple evaluation metrics. Our analysis shows that mechanistic models outperform their agnostic counterparts when tested on independent data and show more consistency to their performance in the training with enhanced interpretability. These findings suggest that using biological constraints in the training process can yield more robust and interpretable gene signatures with high translational potential.