henhaobio
Sign in

bioRxivBioinformaticsPreprint

MtbAIM: AI/ML-guided endpoint-specific prediction of small-molecule activity across drug-susceptible and drug-resistant Mycobacterium tuberculosis

Jasdeep SinghiD, Kritika SinghiD

PDFDOICode

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (12 Oct 2026) · CC BY 4.0

Tree models on RDKit/Morgan features predict M.tb activity and MIC endpoints for drug-susceptible and resistant isolates, slightly beating pretrained embeddings; released as a Colab tool.

Compounds active against drug-susceptible M.tb may not stay active against drug-resistant isolates. The authors say most earlier workflows emphasize H37Rv, pooled MIC values or a single active/inactive label, whereas this one models DS and DR records separately and keeps MIC, MIC90 and MIC99 as independent targets.

Key results

  1. ShownIn scaffold-held-out classification, tree models on RDKit/Morgan features beat the best ChemBERTa neural model for both DS and DR M.tb.

    DS: ExtraTrees balanced accuracy 0.782±0.010, AUROC 0.864±0.007, AUPRC 0.870±0.008 vs ChemBERTa MLP 0.738, 0.813, 0.822. DR: histogram gradient boosting 0.762, 0.836, 0.880 vs ChemBERTa MLP 0.749, 0.830, 0.867. Five-fold means. · Abstract; Results

  2. ShownEndpoint-specific pMIC regression gave useful quantitative separation on held-out scaffolds for all six DS/DR endpoint datasets, using RDKit/Morgan features with tree-based regressors.

    Pooled out-of-fold RMSE/R, DS: MIC 0.714/0.490, MIC90 0.682/0.601, MIC99 0.670/0.541. DR: MIC 0.729/0.546, MIC90 0.784/0.613, MIC99 0.604/0.790. Best models were ExtraTrees (DS MIC, all DR), LightGBM (DS MIC90), XGBoost (DS MIC99). · Abstract; Results

  3. ShownPretrained embeddings (ChemBERTa, MolCLR, Uni-Mol) did not beat classical features in regression; the classical advantage was significant for all three DS endpoints, whereas in DR it was not statistically resolved for MIC90 or MIC99 (the MIC gap was +0.033 pMIC, with significance not stated).

    Best neural model had higher RMSE than best classical model in DS by 0.048 (MIC), 0.069 (MIC90) and 0.056 (MIC99) pMIC, all significant after FDR correction. In DR the gap was +0.033 pMIC for MIC and not resolved for MIC90 or MIC99. · Results; Discussion

  4. SuggestedSmall endpoint datasets give less certain estimates; the DR MIC99 model has the highest R but the least data.

    DR MIC99 pooled R 0.790 on only 180 compounds, vs 3,220 for DR MIC; DS MIC99 has 1,846 compounds vs 21,683 for DS MIC. The authors say the DR MIC99 estimate is less stable than the headline value suggests. · Replicate aggregation and regression targets; Results; Discussion

  5. ShownEight locked models, all on RDKit/Morgan features, were deployed in MtbAIM, a Google Colab tool for DS/DR classification and endpoint-specific potency prediction.

    Accepts SMILES or TXT/CSV lists, up to 100 compounds per submission; returns DS and DR classifications, predicted MIC-endpoint potencies and prediction intervals as downloadable CSV. · Finalized prediction models and MtbAIM platform; Code and model availability

Methods, data and reproducibility

Whole-cell M.tb H37Rv activity records from ChEMBL were split into DS and DR sets, with MIC, MIC90 and MIC99 modeled separately (binary classification and pMIC regression). Classical learners on RDKit/Morgan features were compared with MLPs on ChemBERTa, MolCLR and Uni-Mol embeddings, using frozen five-fold Bemis-Murcko scaffold splits.

  • DS binary classification set from ChEMBL version 37 (12,591 active, 13,054 non-active; active <5,000 nM, non-active >15,000 nM) · Mycobacterium tuberculosis (H37Rv) · n = 25,645 compounds; 8,601 scaffolds
  • DR binary classification set from ChEMBL version 37 (2,215 active, 1,534 non-active) · Mycobacterium tuberculosis (H37Rv) · n = 3,749 compounds; 1,470 scaffolds
  • DS regression datasets (pMIC = -log10 MIC in mol/L), median-aggregated per endpoint · Mycobacterium tuberculosis (H37Rv) · n = 21,683 MIC; 3,754 MIC90; 1,846 MIC99 compounds
  • DR regression datasets (pMIC), median-aggregated per endpoint · Mycobacterium tuberculosis (H37Rv) · n = 3,220 MIC; 561 MIC90; 180 MIC99 compounds
Methods & tools
ChEMBL (version 37), RDKit, Morgan fingerprints, ChemBERTa-77M-MLM, MolCLR, Uni-Mol, Optuna, ExtraTrees, histogram gradient boosting, XGBoost, LightGBM, CatBoost
Reproducibility gaps
No data deposit, curated dataset link, package versions or compute details are given. Raw-data CSVs are mentioned for figures but no location is stated.
Limitations (5)
  • Public bioactivity data are heterogeneous in assay conditions, strain annotation, censoring and measurement density. · stated by the authors
  • Scaffold-aware cross-validation reduces analogue leakage, but prospective temporal or independently generated experimental data are needed to show performance on truly unseen chemical series. · stated by the authors
  • MtbAIM is meant for compound prioritization and does not replace microbiological testing. · stated by the authors
  • The text does not describe how DR records were defined or whether specific resistance profiles were distinguished, so the DR models may mix different resistance contexts. · noted in this summary
  • Compounds with activity between 5,000 and 15,000 nM were excluded from classification, so classifier metrics reflect clear-cut actives and non-actives only. · noted in this summary

Read this if you want scaffold-split benchmarks of classical vs pretrained molecular representations for antitubercular activity, split by drug-susceptible vs drug-resistant M.tb and by MIC endpoint.

Citations
0

Abstract

The emergence of drug resistance in Mycobacterium tuberculosis (M.tb) presents a major challenge to drug discovery, as compounds active against drug-susceptible (DS) isolates may not retain comparable inhibitory activity against drug-resistant (DR) isolates. Here, we developed an AI/ML workflow to independently prioritize small molecules separately against DS and DR isolates of M.tb and predict MIC, MIC90, and MIC99 as distinct quantitative endpoints. Following curation of data retrieved from the ChEMBL repository, molecular-activity datasets comprised 25,645 and 3,749 compounds tested against whole cell DS and DR isolates respectively. Molecular representations were generated using RDKit descriptors and Morgan fingerprints, along with embeddings from ChemBERTa, MolCLR, and Uni-Mol, and evaluated for binary activity classification and endpoint-specific regression. The best-performing DS classifier combined RDKit/Morgan features with ExtraTrees (balanced accuracy 0.782; AUROC 0.864; AUPRC 0.870), while histogram gradient boosting using the same features performed best for DR classification (balanced accuracy 0.762; AUROC 0.836; AUPRC 0.880). Similarly, RDKit/Morgan features combined with tree-based regressors achieved the best overall performance across all six DS/DR regression endpoints, with pooled out-of-fold RMSE values ranging from 0.604 to 0.784 pMIC units. To facilitate broader application, we developed MtbAIM (https://github.com/jasdeep002/MtbAIM), a Google Colab-based prediction platform that enables users to assess and prioritize small molecules against DS and DR isolates of M.tb.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLiu '26Zabeti '21Lv '25consortium '21Alderwick '18Serajian '24
72% semantic matchBioinformatics
C X Liu, Haoyue Zhu, Pengfei Zhou+9 more·bioRxiv·Jul 3, 2026