Enhancing the Predictive Power of Macrocyclic Drug Permeability by Knowledge Distillation from Analogous Pretraining...

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Pharmacology

Background:

• Macrocyclic drugs are potent modulators of protein-protein interactions.
• Poor and unpredictable membrane permeability hinders macrocycle drug development.
• Current experimental and 3D modeling methods are slow and computationally intensive.

Purpose of the Study:

• To develop a deep learning (DL) model for predicting macrocycle permeability directly from 2D structures.
• To overcome the limitations of experimental testing and 3D modeling in macrocycle drug discovery.
• To enable efficient prioritization of macrocycles with favorable pharmacokinetic properties.

Main Methods:

• Developed Multi_DDPP, a deep learning model predicting macrocycle permeability from 2D structures.
• Employed knowledge distillation using multi-cell line permeability data for improved generalizability.
• Integrated diverse molecular representations (physicochemical descriptors, fingerprints, molecular graphs, hybrid features).
• Utilized node masking for substructure identification and regression extensions for physiological parameter incorporation.

Main Results:

• Multi_DDPP outperforms existing machine learning (ML) and DL approaches in permeability prediction.
• The model effectively leverages diverse molecular representations and knowledge distillation.
• Node masking identifies key substructures influencing permeability.
• Regression extensions refine predictions with physiological parameters.

Conclusions:

• Multi_DDPP enables rapid, 2D-based permeability prediction, bypassing costly 3D conformer generation.
• The model facilitates efficient screening and prioritization of macrocyclic drug candidates.
• This approach significantly enhances the early-stage drug discovery process for macrocycles.