MycoPermeNet-v2: Improved Prediction of Mycomembrane Permeation Using Fusion Noisy Student Self-Distillation.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Tuberculosis (TB) drug discovery faces challenges due to the intrinsic antibiotic resistance of *Mycobacterium tuberculosis*.
• The mycomembrane of *M. tuberculosis* acts as a permeability barrier, hindering antibiotic efficacy.
• Existing machine learning models struggle to generalize for predicting compound permeability due to limited labeled data.

Purpose of the Study:

• To develop a robust machine learning model for predicting compound permeability across the *M. tuberculosis* mycomembrane.
• To address data scarcity issues in antibiotic discovery using advanced machine learning techniques.

Main Methods:

• Proposed a two-stage model, MycoPermeNet-v2, integrating molecular descriptors with graph-based embeddings.
• Employed Noisy Student self-distillation (NST) to enhance model performance with limited labeled data.
• Conducted systematic evaluations of model robustness, component contributions, and interpretability.

Main Results:

• Achieved significantly improved prediction performance (RMSE reduced from 0.755 ± 0.024 to 0.719 ± 0.021, adjusted *p* < 0.0001).
• Demonstrated that the model captures chemically meaningful features relevant to mycomembrane permeability.
• Showcased generalizability across different models and physicochemical property prediction tasks.

Conclusions:

• MycoPermeNet-v2 offers a robust solution for predicting compound permeability in data-constrained scenarios.
• The model's interpretability provides insights into key features for *M. tuberculosis* drug permeability.
• This approach holds strong applicability for accelerating antibiotic discovery, particularly for challenging pathogens like *M. tuberculosis*.