Interpretable detection of novel human viruses from genome sequencing data

Jakub M Bartoszewicz; Anja Seidel; Bernhard Y Renard

In the Fig. 1 we present example filters, visualized as "maxcontrib" sequence logos based on mean partial Shapley values for each nucelotide at each position. All nucleotides of the filters with the highest (Fig. 1a) or lowest (Fig. 1b) score have relatively strong contributions in accordance with the filters' own contributions. However, we observe that sometimes, there is a "conserved" nucleotide which consistently appears in the activating subsequences, but the sign of its contributions is opposite to the filter's (Fig. 1c) . Those "counter-contributions" may arise if a nucleotide with a negative weight forms a frequent motif with others with positive weights strong enough to activate the filter. We comment on this fact in the Section 4.2. Some filters seem to learn gapped motifs resembling a codon structure (Fig. 1d) . We extracted this filter from a network predicting bacterial pathogenicity trained by Bartoszewicz et al. (2019) , but we find similar filters in our networks as well. We scanned a genome of S. aureus with this filter and discovered that the learned motif is indeed significantly enriched in coding sequences (Fisher exact test with Benjamini-Hochberg correction, q < 10 −15 ). It is also enriched in a number of specific genes. The one with the most hits (sraP, q < 10 −15 ) is associated with virulence in endovascular infection.