Fig. 2: Models employed to predict depolymerase-dependent phage tropism.

The elements of the architectures are described using as an illustration the scenario of a lysogen with two prophages, one carrying two de polymerase sequences and the other carrying one 1. A Description of the architecture of TropiGAT. The input is modelized with graphs, where the bacteria, prophage and the encoded depolymerase are represented as nodes and their relationship by edges. The features of the depolymerase are aggregated by the encoder before being classified. B Description of the architecture of TropiSEQ. The depolymerase domain sequences are retrieved and then used to scan the collection of representative depolymerase domain clusters using BLASTp. The vector representing the presence-absence of each of the depolymerase domain cluster is fed into a Random Forest classifier. C TropiGAT’s (n = 85KL types) and D TropiSEQ’s (n = 108KL types) scatter plots representing the MCC scores on the y-axis, along with the number of infecting prophages on the x-axis, used for training. The green line in bold represents the regression line, and the colored area the deviation. r coefficient = Pearson correlation coefficient.