Deep Learning for Predicting 16S rRNA Gene Copy Number

Abstract

ABSTRACT16S rRNA is a commonly used technique to identify prokaryotes. However, this approach could only reflect the proportion of sequencing reads, instead of actual cell fraction. To accurately represent actual cell fraction, bioinformatic tools that estimate 16S gene copy number (GCN) of unmeasured species are required. All the existing tools rely on taxonomy or phylogeny to predict 16S GCN. The present study develops a deep learning-based method that establishes a novel route. The proposed approach, i.e., Artificial Neural Network Approximator for 16S rRNA Gene Copy Number (ANNA16), essentially links 16S sequence string directly to GCN, without the construction of taxonomy or phylogeny. The computational analysis revealed the success of the deep learning model built on 16S GCN data retrieved from rrnDB (N = 19,520). The resulting deep learning model outperforms all the major and state-of-the-art algorithms in the literature. Shapley Additive exPlanations (SHAP) shows ANNA16 is capable of detecting informative positions and weighing K-mers unequally according to their informativeness to more effectively utilize the information contained in 16S sequence. For the accessibility of ANNA16, it will be released as a public, an end-to-end, web-based tool that predicts copy number from 16S rRNA sequence.