Interpretable Predictive Modeling of Host Specificity of Rabies lyssavirus from Glycoprotein Sequences

Abstract Rabies lyssavirus (RABV) is a zoonotic virus that affects a wide range of mammalian hosts and has a 100% mortality rate following the onset of symptoms. Let s = (a1 , . . . , aL ) denote the amino acid sequence of the RABV G protein and let y ∈ Y denote the host isolate. This study learns a family of parameterized mapping functions f θ : X → ∆|Y| that map an encoded sequence x = φ(s) ∈ X to a predictive host probability vector pˆ = f θ (x) . Among the learned mappings, the long short-term memory (LSTM) mapping composed with an integer-encoding operator achieved the highest test accuracy of 90.0%. Additionally, interpretability analyses of fθ recover known antigenic sites and identify a strong association between residue positions 460–520 and host specificity. Overall, the learned maps and their sensitivity analyses provide groundwork toward targeted vaccines and containment.