Explaining the pathogenesis of African swine fever using knowledge-driven regulatory network modeling

A lethal DNA virus with significant economic impact on livestock farmers worldwide, the pathogenesis of African swine fever virus (ASFV) infection is complex and continues to challenge the development of effective vaccine candidates. The requirement for high-containment conditions further complicates its study, resulting in limitations in sample size and marker assessment that challenge conventional statistical analysis. In this work we demonstrate how prior knowledge of immune biology and pathogen-host proteome interactions can be leveraged and reconciled with sparse experimental data to deliver plausible mechanistically informed hypotheses describing ASFV illness progression. We apply large-scale automated mining of literature and pathway schema together with generative artificial intelligence (AI) to create closed-loop regulatory network models consisting of 133 pathogen and host proteins linked by 676 regulatory interactions. Immune regulatory tuning of these networks is reverse engineered to explain two distinct experimentally observed illness progression trajectories in only 5 markers measured every second day over a maximum of 8 days. Comparison of network model pools specific to each progression phenotype suggest that these significantly different outcomes may arise from altered regulatory tuning of genes coding for interleukin (IL)1{beta}, tumor necrosis factor (TNF) and Forkhead box protein (FOX)O4, potentially as a result of epigenetic adaptations. Simulated challenges with individual ASFV protein confirm broadly delayed interferon (IFN)-{gamma}I response in both phenotypes, with multigene family (MGF)505-3R offering the earliest induction and only in the more severe phenotype. Paradoxically, predictions suggest that this delay is preceded by an early IL-10 induction by this same viral protein. While added model granularity and validation is needed, we propose that this proof-of-concept knowledge driven approach offers an attractive solution to mechanistic hypothesis generation in data poor environments.