Machine Learning and Chemical Features Unlock Mycomembrane Permeation in Mycobacterium tuberculosis

A multidisciplinary team led by Irene Lepori and M. Sloan Siegrist has mapped the chemical features that govern permeation across the notoriously impermeable mycomembrane of Mycobacterium tuberculosis (搜索)搜索该疾病, combining high-throughput screening with machine learning to predict and engineer improved cell-envelope penetration. The work, published in Nature Microbiology, addresses a central challenge in antitubercular drug development: designing molecules that can cross the bacterial cell envelope, which is often more difficult than achieving target binding.

A High-Throughput Assay for Mycomembrane Permeability

The researchers employed a Peptidoglycan Accessibility Click-Mediated Assessment (PAC-MAN) assay to systematically measure mycomembrane permeation. In this approach, the mycobacterial peptidoglycan is first metabolically labelled with the strained alkyne dibenzocyclooctyne (DBCO). Bacteria are then incubated with azide-functionalized test molecules; compounds that successfully permeate the mycomembrane react with the DBCO through strain-promoted azide–alkyne cycloaddition (SPAAC). The remaining unreacted DBCO groups are subsequently labelled with an azide-functionalized fluorophore, so that fluorescence intensity serves as an inverse readout of mycomembrane permeation.

Using PAC-MAN, the team screened M. tuberculosis mc26206 (H37Rv ΔpanCD ΔleuCD) and the model organism M. smegmatis (Msm) mc2155 with 1,572 azide-test molecules drawn from three sources: 1,152 synthesized via fluorosulfuryl azide chemistry, 380 purchased from Enamine, and 40 from various commercial sources. To control for the intrinsic reactivities of azide-test molecules toward DBCO, the researchers also screened DBCO-functionalized polystyrene beads that have bacteria-like dimensions but lack a permeability barrier. Log-linear regression analyses yielded standardized residuals that were consistent across libraries.

Chemical Features Associated with Permeation

Cheminformatics analysis revealed that aromatic nitrogen-containing scaffolds such as indole, imidazole, and pyrazole correlate positively with mycomembrane permeation, while scaffolds such as cyclopentane or cyclohexane correlate negatively. Notably, the physicochemical properties previously shown to correlate with Gram-negative accumulation were not associated with mycomembrane permeation.

When analysed as a whole, the datasets revealed weak or no correlations between physicochemical properties and permeation. However, after grouping compounds by scaffold, clear correlations emerged. For example, topological polar surface area (TPSA) showed a strong negative correlation with permeation when a molecule contains an indazole, while log partition coefficient (logP) showed a strong negative correlation when a molecule contains naphthalene. The authors note these observations are "striking as lipophilicity is generally viewed as a positive attribute for antitubercular drugs."

A Machine Learning Model to Predict Permeation

To capture the complex relationships between chemical structure and mycomembrane permeability, the team built a machine learning model called Mycobacterial Permeability neural Network (MycoPermeNet (搜索)搜索该靶点). The model takes Simplified Molecular Input Line Entry System (SMILES) strings and Mtb screening data as inputs, using a two-stage deep-learning process: it first generates vector representations (embeddings) of compounds, then uses a multilayer perceptron to convert embeddings into permeability predictions.

MycoPermeNet (搜索)搜索该靶点 was trained on 80% of the permeability data, with 10% used for validation and 10% as a test set selected using the Bemis–Murcko scaffold split. The final model achieved a Spearman rank correlation coefficient (ρ) of 0.81 ± 0.019 on the train set and 0.74 ± 0.020 on the test set, indicating that the model "correctly ranks the relative permeability of compounds even better than it predicts their absolute permeability scores."

Among the 20 scaffolds predicted as most permeable, the model identified various indole-, imidazole-, and pyrazole-like scaffolds, concordant with the cheminformatics analysis. Interpretability studies using a surrogate XGBoost algorithm and Shapley Additive exPlanations (SHAP) identified TPSA and logP as the two most influential physicochemical properties driving predictions.

Experimental Validation of Causative Features

The researchers tested whether the identified molecular features are causative by synthesizing small-molecule series. In a series based on JSF-2985 (搜索)搜索该药品, an antitubercular molecule previously reported by the group, derivatives bearing imidazole, pyrazole, and pyrrolidine scaffolds (Δlog10CC50 = 1.1–1.3) permeated the mycomembrane better than a derivative bearing cyclopentane (Δlog10CC50 = 2), consistent with both cheminformatics and MycoPermeNet (搜索)搜索该靶点 predictions.

A second series based on a pentapeptide (Phe-Lys-Phe-Lys-Phe) systematically substituted phenylalanines for tryptophans ("W peptides"), swapping benzene for indole. Consistent with predictions, these substitutions enhanced mycomembrane permeation in both Mtb and Msm, and also resulted in higher overall Msm cell accumulation as measured by LC–MS.

The benzene-to-indole substitution proved generalizable in a third series based on octyl tridecaptin A1 (搜索)搜索该药品 (OctTriA1), an antimicrobial peptide. Replacing the phenylalanine at position 9 with tryptophan generated OctTriA5. Across four matched pairs, methylated azide-OctTriA5 permeated the Mtb mycomembrane better than methylated azide-OctTriA1.

Indole Association with Whole-Cell Activity

The researchers next asked whether modulating mycomembrane permeation impacts whole-cell anti-Mtb activity. For the JSF-2985 (搜索)搜索该药品 analogues, Mtb growth inhibition did not reflect permeation profiles, and a sub-inhibitory dose of ethambutol查看药品详情—an antibiotic that indirectly disrupts the mycomembrane—did not potentiate JSF-2985 activity. The authors suggest the mycomembrane may not be an important barrier to JSF-2985's whole-cell activity, or that the molecule's small size (309 Da) complicated scaffold modifications without impacting target engagement.

By contrast, OctTriA1 (1,521 Da) was sensitized by sub-MIC ethambutol查看药品详情, and both OctTriA1 and OctTriA5 were shown to use mycobacterial lipid II (搜索)搜索该靶点 as a receptor, inhibiting Msm growth via a mechanism distinct from vancomycin查看药品详情. In the four matched pairs of OctTriA1 and OctTriA5 derivatives, OctTriA5 analogues better inhibited the growth of both Msm and Mtb.

A retrospective analysis of the ~200,000-compound Molecular Libraries Small Molecule Repository (MLSMR) further supported the findings: in two distinct molecule sets differing in the presence or absence of indole in peripheral positions, compound activity correlated with the presence of indole and with ML-predicted mycomembrane permeation.

Permeability Predictors Correlate with Whole-Cell Activity

Finally, the researchers examined whether permeability predictors correlate with whole-cell activity in large datasets. Across three screens—the MLSMR and Tuberculosis Antimicrobial Acquisition and Coordinating Facility (TAACF) collections screened against whole Mtb cells, and a third screen against the purified Mtb enzyme Rv3671c—molecule activity in whole-cell screens correlated with both observed scaffold permeability and MycoPermeNet (搜索)搜索该靶点 predictions. Critically, these correlations were absent in the enzyme screen, a negative control for the mycomembrane barrier.

"These data suggest that scaffolds and other chemical features that influence mycomembrane permeation also predict whole-cell anti-Mtb activity in large datasets," the authors conclude. The work establishes a framework for designing antitubercular compounds that can overcome one of the pathogen's most formidable defenses, with potential implications for accelerating drug discovery against Mycobacterium tuberculosis (搜索)搜索该疾病.