Fig. 1

Overview of HMD-ARG. Top panel: HMD-ARG is composed of three deep learning models, which are responsible for three level predictions. In level 0, one model is trained to predict whether an input sequence is an ARG or not. If it is an ARG, it will go through the second level prediction, in which a multi-task deep learning model (more details shown in the bottom panel) is trained to predict the resistant antibiotic family, resistant mechanism, and gene mobility information at the same time. If the sequence is predicted as beta-lactamase in level 1, it will be fed into the level 2 model to predict its beta-lactamase subclass. Bottom panel: In order to train those models, we built the most comprehensive ARG database to date by merging the sequences from seven existing databases, followed by a post-processing step to remove duplicates. Then, we used the existing tools and manual curation to assign the annotations from three aspects, i.e., resistant antibiotic family, resistant mechanism, and gene mobility, to each sequence in the database. Those sequences were then fed to deep learning models to train our models, as illustrated in the right part