Predicting Viral Spillover Risk using Machine Learning and Adversarial Mutation

Project Summary The potential for animal viruses to spill over into humans represents a persistent, consequential threat to human global health. However, our ability to predict which viruses are most likely to “jump” into humans is currently limited. Computational approaches are a potential solution to this problem, and early results are promising. Nevertheless, most computational approaches are “static” with respect to virus evolution - they endeavor to predict zoonotic potential of viruses at a single point in time (the point at which viruses were identified or sequenced), but they do not account for evolution, which is a hallmark of virus biology. This project will fill this methodological gap by offering a new approach inspired by adversarial machine learning. Adversarial machine learning is a broad area of research that includes methods designed to identify small changes to a learned model’s input that significantly change its output. We will develop and apply adversarial machine learning to the following prediction task: given genomic protein sequences from a non- human animal virus, use a learned model to predict if the virus is likely to be infectious to humans, or how extensively it would have to mutate to become infectious Our proposed approach will notably advance the state of the art by (i) considering not just the human-infection risk for a “static” viral sequence but also the spillover risk attributable to evolutionary variants, (ii) applying and developing methods to characterize and explain the risk prediction for a given virus, and (iii) developing and evaluating predictive models based on state-of-the-art neural networks for protein sequences.