Artificial Intelligence and Machine learning based prediction of resistant and susceptible mutations in Mycobacterium tuberculosis

Abstract

Tuberculosis (TB), an infectious disease caused by Mycobacterium tuberculosis (M.tb), causes highest number of deaths globally for any bacterial disease necessitating novel diagnosis and treatment strategies. High-throughput sequencing methods generate a large amount of data which could be exploited in determining multi-drug resistant (MDR-TB) associated mutations. The present work is a computational framework that uses artificial intelligence (AI) based machine learning (ML) approaches for predicting resistance in the genes rpoB, inhA, katG, pncA, gyrA and gyrB for the drugs rifampicin, isoniazid, pyrazinamide and fluoroquinolones. The single nucleotide variations were represented by several sequence and structural features that indicate the influence of mutations on the target protein coded by each gene. We used ML algorithms - naïve bayes, k nearest neighbor, support vector machine, and artificial neural network, to build the prediction models. The classification models had an average accuracy of 85% across all examined genes and were evaluated on an external unseen dataset to demonstrate their application. Further, molecular docking and molecular dynamics simulations were performed for wild type and predicted resistance causing mutant protein and anti-TB drug complexes to study their impact on the conformation of proteins to confirm the observed phenotype.

Figure 1

ROC plots for all the models generated for genes (A) rpoB, (B) pncA, (C) inhA, (D) katG, (E) gyrA and (F) gyrB.

Figure 2

RMSD, Rg and SASA plot for pncA gene. The RMSD, Rg and SASA were less in case of wild type indicating that the mutations destabilized the protein.

Figure 3

RMSD, Rg and SASA plot for katG gene. The RMSD, Rg and SASA of mutants were higher that wild type demonstrating that the wild type protein was more stable.

Figure 4

Interaction patterns between (A) wild type and (B) L587I (C) L619P (D) L634F (E) N238K mutant protein-isoniazid complexes. The drug bound to protein through hydrophobic interactions only, however strong binding was observed in wild type protein.

Figure 5

Hydrogen bonding and hydrophobic interactions seen in (A) wild type, (B) L96E and (C) V155M mutant protein-pyrazinamide complexes. Fewer interacting residues were observed in case of mutants in comparison to wild type.

Figure 6

RMSD, Rg and SASA plot for gyrA gene, N-terminal protein. The plots for RMSD, Rg and SASA were similar to wild type in case of mutant, L711M.

Figure 7

RMSD, Rg and SASA plot for gyrA gene, C-terminal protein. For Q431E mutant, the RMSD and Rg were slightly higher than wild type, however SASA was less for mutant protein.

Figure 8

Interaction pattern observed between N-terminal of wild type gyrase A and fluoroquinolones; (A) ofloxacin; (B) moxifloxacin; (C) ciprofloxacin and mutant, L711M; (D) ofloxacin; (E) moxifloxacin and (F) ciprofloxacin. The wild type protein formed hydrogen bonds with the drugs whereas no hydrogen bond was present in case of mutant protein-drug complexes.

Figure 9

Interaction pattern observed between C-terminal of wild type gyrase A and fluoroquinolones; (A) ofloxacin; (B) moxifloxacin; (C) ciprofloxacin and mutant, Q431E; (D) ofloxacin; (E) moxifloxacin and (F) ciprofloxacin. More number of interacting residues were present in wild type protein bound to the drugs than in mutant protein-drug complexes.

Figure 10

RMSD, Rg and SASA plot for gyrB gene. The RMSD was higher for mutant while Rg and SASA were approximately similar for wild type and mutant showing that mutation did not had much impact on the protein.

Figure 11

Hydrogen bonding and hydrophobic interactions between wild type gyrase B and various drugs (A) ofloxacin; (B) moxifloxacin; (C) ciprofloxacin and mutant protein, N499T; (D) ofloxacin (E) moxifloxacin and (F) ciprofloxacin. In case of mutant proteins, only weak hydrophobic interactions were seen.