# Evaluation of Marker Gene-Based In Silico Antimicrobial Resistance Prediction Tools

## TL;DR

This study evaluates the accuracy of in silico tools for predicting antimicrobial resistance using 16S rRNA data and finds that specialized databases and improved algorithms are needed for better results.

## Contribution

The study compares the performance of three 16S rRNA-based AMR prediction tools and highlights the need for improved methods and databases.

## Key findings

- 16S rRNA-based AMR prediction tools showed low F1 scores, with Tax4Fun performing best at 0.22.
- Specialized AMR databases and improved algorithms are needed for meaningful resistance prediction accuracy.
- Shotgun metagenomics outperformed 16S rRNA-based methods in AMR profiling.

## Abstract

While 16S rRNA-based predictive functional profiling has proven useful for broad ecological studies, its utility for AMR surveillance remains limited. The results of our study highlight the necessity of integrating specialized AMR databases and improving algorithmic approaches to achieve meaningful accuracy in resistance prediction. These advancements will be essential if marker gene-based tools are to complement or substitute for shotgun metagenomics in the context of clinical or epidemiological AMR monitoring.

The monitoring and surveillance of antimicrobial resistance (AMR) is an important procedure in clinical patient management and epidemiological public health. Conventionally, culture-based tools such as disk diffusion methods or broth dilution methods for antibiotic susceptibility tests are used. While culture-independent approaches, such as PICRUSt2, Tax4Fun, or MicFunPred, have recently been tried based on predictive functional profiling using the 16S rRNA marker gene, evaluations of AMR tools are scarce. A total of 20 E. coli strains (Carbapenem-resistant (CRE) positive: 10, CRE negative: 10) were used. The AMR phenotype was based on Vitek2 (bioMerieux). DNA was extracted from the 20 strains, and 16S rRNA (V3-V4 region) and shotgun sequencing was carried out. The bioinformatic pipelines were QIIM2 for 16S rRNA and MetaPhlAn4 for shotgun. The functional prediction tools were PICRUSt2, Tax4Fun, and MicFunPred for 16S rRNA and AMRFinderPlus for shotgun. The presence/absence of 23 KEGG numbers regarding AMR in PICRUSt2, Tax4Fun, and MicFunPred were compared to shotgun AMR profiles. The F1 scores were calculated according to each 16S marker gene-based prediction tool using a confusion matrix. A total of 12 classes of antibiotics, including carbapenem, were analyzed. The F1 scores of 16S predictive functional profilers regarding AMR were 0.22 for Tax4Fun, 0.12 for PICRUSt2, and 0.08 for MicFunPred. While Tax4Fun showed the highest F1 score of the three 16S predictive functional profilers, the F1 scores were generally low. Our study highlights the necessity of integrating specialized AMR databases and improving algorithmic approaches to achieve meaningful accuracy in resistance prediction.