Pretrained graph transformer-driven de novo design and machine learning prioritization of computationally predicted ENR-InhA (InhA) inhibitors with potential relevance to drug-resistant Mycobacterium tuberculosis

Multidrug-resistant tuberculosis (MDR-TB) undermines frontline therapy, and conventional drug discovery is slow and resource intensive. We present an integrated computational pipeline for the de novo design of inhibitors of a single clinically validated target, the NADH-dependent enoyl-[acyl-carrier-protein] reductase of Mycobacterium tuberculosis (ENR-InhA, encoded by inhA). Bioactive compounds were retrieved from ChEMBL and curated to 80 with equality-qualified IC50 values. Four conventional machine-learning models and one deep neural network were trained on Morgan and MACCS fingerprints to predict pIC50. DrugGEN, a pretrained graph-transformer generative adversarial network, designed de novo molecules. Of 74 generated, 67 were novel relative to the curated set (nearest-neighbour Tanimoto < 0.85), and the 23 with predicted pIC50 ≥ 7.0 proceeded to ADMET screening and docking. Three compounds met all drug-likeness criteria, raised no toxicity alerts in the models applied, and gave docking scores (−8.27, −7.66, and −7.76 kcal/mol) more negative than direct ENR-InhA inhibitors docked identically. All three complexes maintained stable RMSD profiles over 200 ns of molecular dynamics, and compound 3 showed the greatest rigidity and broadest interaction profile. These are computationally prioritized candidates of potential relevance to drug-resistant tuberculosis; none was tested against M. tuberculosis or drug-resistant isolates.