GenTB: A user-friendly genome-based predictor for tuberculosis resistance powered by machine learning

We present Translational Genomics platform for Tuberculosis (GenTB), a free and open web-based application to predict antibiotic resistance from next-generation sequence data. The user can choose between two potential predictors, a Random Forest (RF) classifier and a Wide and Deep Neural Network (WDNN) to predict phenotypic resistance to 13 and 10 anti-tuberculosis drugs, respectively. We benchmark GenTB's predictive performance along with leading TB resistance prediction tools (Mykrobe and TB-Profiler) using a ground truth dataset of 20,408 isolates with laboratory-based drug susceptibility data. All four tools reliably predicted resistance to first-line tuberculosis drugs but had varying performance for second-line drugs. The mean sensitivities for GenTB-RF and GenTB-WDNN across the nine shared drugs were 77.6% (95% CI 76.6-78.5%) and 75.4% (95% CI 74.5-76.4%), respectively, and marginally higher than the sensitivities of TB-Profiler at 74.4% (95% CI 73.4-75.3%) and Mykrobe at 71.9% (95% CI 70.9-72.9%). The higher sensitivities were at an expense of ≤ 1.5% lower specificity: Mykrobe 97.6% (95% CI 97.5-97.7%), TB-Profiler 96.9% (95% CI 96.7 to 97.0%), GenTB-WDNN 96.2% (95% CI 96.0 to 96.4%), and GenTB-RF 96.1% (95% CI 96.0 to 96.3%). Averaged across the four tools, genotypic resistance sensitivity was 11% and 9% lower for isoniazid and rifampicin respectively, on isolates sequenced at low depth (< 10× across 95% of the genome) emphasizing the need to quality control input sequence data before prediction. We discuss differences between tools in reporting results to the user including variants underlying the resistance calls and any novel or indeterminate variants CONCLUSIONS: GenTB is an easy-to-use online tool to rapidly and accurately predict resistance to anti-tuberculosis drugs. GenTB can be accessed online at https://gentb.hms.harvard.edu , and the source code is available at https://github.com/farhat-lab/gentb-site .