Multi-modal machine learning and molecular modelling reveal structurally diverse inhibitors of Mycobacterium tuberculosis protein tyrosine phosphatase B

Protein tyrosine phosphatase B (PtpB) is a virulence-associated phosphotyrosine phosphatase secreted by Mycobacterium tuberculosis (Mtb), known to disrupt host immune signaling by dephosphorylating key proteins. Targeting PtpB represents a rational strategy for anti-TB drug discovery. This study presents an integrative computational framework for identifying and evaluating small-molecule inhibitors of Mtb PtpB. QSAR models were constructed using four molecular fingerprint types, CDK, PubChem, MACCS, and AtomPairs2DCount, as regression models predicting pIC 50 values. Multiple machine learning algorithms were evaluated, with model performance assessed via R 2 , RMSE, cross-validation, and Y-randomization. SHAP analysis was applied to the top-performing PubChem-SVR model to interpret key structural features. Top-ranked compounds were subjected to molecular docking followed by 250 ns MD simulations to examine binding stability. MM-GBSA and PCA were used for post-simulation analysis. Gene-level interactions were evaluated by comparing predicted compound targets with Mtb-related host genes. Among descriptors, the PubChem-RF model achieved the best performance. SHAP identified PubchemFP417 (alkyne), PubchemFP462 (carboxylic acid), PubchemFP143 (five-membered rings), and PubchemFP34 (sulfur-containing fragments) as major contributors. CHEMBL4635765 showed strong and stable binding within the PTP pocket, while isoxazole carboxylic acid maintained key interactions but with lower stability. Network analysis revealed four shared targets (APP, HDAC8, CACNA1B, pvdQ) and compound-specific links to immune-related genes, including PTPN1 and NFKB1. This integrative computational study combines machine learning, structural modeling, and network pharmacology to provide mechanistic insights into PtpB inhibition and to identify promising chemical scaffolds for future anti-tubercular research. As the analysis is entirely computational, experimental validation will be required to confirm the predicted activities.