Predictive Modeling of PROTAC Cell Permeability with Machine Learning.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Predicting cell permeability is crucial for developing effective proteolysis targeting chimeras (PROTACs), reducing costly synthesis and testing.
• Existing methods for PROTAC permeability prediction require improvement to accelerate the drug discovery pipeline.

Purpose of the Study:

• To investigate the effectiveness and limitations of machine learning (ML) models for predicting cell permeability in cereblon (CRBN) and von Hippel-Lindau (VHL) PROTACs.
• To identify key molecular descriptors influencing PROTAC cell permeability.

Main Methods:

• Development of binary classification ML models using 17 simple descriptors for large, structurally diverse sets of CRBN and VHL PROTACs.
• Evaluation of model performance using kappa nearest neighbor and random forest algorithms on blinded test sets.
• Analysis of descriptor importance for model accuracy.

Main Results:

• For VHL PROTACs, kappa nearest neighbor and random forest models achieved >80% accuracy (κ ≥ 0.57) in predicting permeability for blinded test sets.
• Retraining models with combined training and test data maintained high performance for VHL PROTACs.
• CRBN PROTAC models showed lower success, attributed to imbalanced datasets; size and lipophilicity emerged as key predictive descriptors.

Conclusions:

• Machine learning models, particularly for VHL PROTACs, can be effectively trained to predict cell permeability with high accuracy.
• Properly trained ML models serve as valuable filters in the PROTAC design process, optimizing resource allocation.
• Addressing dataset imbalance is critical for improving ML model performance in PROTAC permeability prediction, especially for CRBN PROTACs.