Interpretable detection of novel human viruses from genome sequencing data

Jakub M Bartoszewicz; Anja Seidel; Bernhard Y Renard

However, some of the training sequences contain N s themselves. It is therefore possible that a filter will learn only negative weights at a given position, even though there is no biological justification for that. This may lead to assigning only negative contributions to all four possible nucleotides at a given position if the filter's contribution is positive (and positive nucleotide contributions if the filter's contribution is negative). To solve the problem, we first normalize the weight matrices position-wise, as described by Shrikumar et al. (2019a) . Finally, we calculate average filter contributions to obtain a crude ranking of feature importance with regard to both the positive and negative class.