DeepSEA: an alignment-free deep learning tool for functional annotation of antimicrobial resistance proteins

Abstract

AbstractSurveying antimicrobial resistance (AMR) is essential to track its evolution and spread. Alignment-based annotation tools use strict identity (>80%) cutoffs to distinguish between non-resistant (NRP) and resistant proteins (ARP) only annotating proteins similar to those in their databases. Deep learning and Hidden Markov Models (HMM) based tools also depend on protein alignment at some level. DeepARG filters input data to select the um SNP ARG-like proteins and HMMs are built on multi-sequence alignment (MSA) specific for the protein in a given family or group. Therefore, there is a need to remove the alignment dependency of AMR annotation tools to identify proteins with remote homology Here we present DeepSEA, an alignment-free tool fitted on antimicrobial-resistant sets of aligned and unaligned ARPs and NRP. DeepSEA outperforms the current multi-class AMR classifiers DeepARG, RGI and AMRfinder. Furthermore, DeepSEA trained weights cluster AMR by resistant mechanisms, indicating that the model’s latent variables successfully captured distinguishing features of antibiotic resistance. Our tool annotated functionally validated tetracycline destructases (TDases) and confirmed the identification of a novel TDase found by HMM.