Viral entry is a critical step in the infection process. Klebsiella spp. and other clinically relevant bacteria often express a complex polysaccharide capsule that acts as a barrier to phage entry. In turn, most Klebsiella phages encode depolymerases for capsule removal. This virus-host arms race has led to extensive genetic diversity in both capsules and depolymerases, complicating our ability to understand their interaction. This study exploits the information encoded in Klebsiella prophages to model the interplay between the bacteria, the prophages, and their depolymerases, using a graph neural network and a sequence clustering-based method. Both approaches showed significant predictive ability for prophages capsular tropism and, importantly, were transferrable to lytic phages. In addition to creating a comprehensive database linking depolymerase sequences to their specific targets, this study demonstrates the predictability of phage-host interactions at the subspecies level, providing new insights for improving the therapeutic and industrial applicability of phages.