Workflow from Scientific Research

Given a compound with a defined activity toward EV71, we represent a compound as the latent code of a variational autoencoder (VAE), a Morgan fingerprint, or their feature concatenation. The dimension of the compound representation is then reduced by principal-component analysis. The reduced-dimension compound feature is coupled with pH value and input into various discriminative machine learning models to predict its probability of having a log reduction >1.5 (i.e., a classification setting) and its exact log reduction value (i.e., a regression setting). Best feature and model are selected based on leave-one-compound-out cross-validation.

DOI