Deep Learning for Predicting 16S rRNA Gene Copy Number

BackgroundCulture-independent 16S rRNA gene metabarcoding is a commonly used method in microbiome profiling. However, this approach can only reflect the proportion of sequencing reads, rather than the actual cell fraction. To achieve more quantitative cell fraction estimates, we need to resolve the 16S gene copy numbers (GCN) for different community members. Currently, there are several bioinformatic tools available to estimate 16S GCN, either based on taxonomy assignment or phylogeny. MethodHere we develop a novel algorithm, Stacked Ensemble Model (SEM), that estimates 16S GCN directly from the 16S rRNA gene sequence strings, without resolving taxonomy or phylogeny. For accessibility, we developed a public, end-to-end, web-based tool based on the SEM model, named Artificial Neural Network Approximator for 16S rRNA Gene Copy Number (ANNA16). ResultsBased on 27,579 16S rRNA gene sequence data (rrnDB database), we show that ANNA16 outperforms the most commonly used 16S GCN prediction algorithms. The prediction error range in the 5-fold cross validation of SEM is completely lower than all other algorithms for the 16S full-length sequence and partially lower at 16S subregions. The final test and a mock community test indicate ANNA16 is more accurate than all currently available tools (i.e., rrnDB, CopyRighter, PICRUSt2, & PAPRICA). SHAP value analysis indicates ANNA16 mainly learns information from rare insertions. ConclusionANNA16 represents a deep learning based 16S GCN prediction tool. Compared to the traditional GCN prediction tools, ANNA16 has a simple structure, faster inference speed without precomputing, and higher accuracy. With increased 16S GCN data in the database, future studies could improve the prediction errors for rare, high-GCN taxa due to current under sampling.