Systematic benchmarking of 13 AI methods for predicting cyclic peptide membrane permeability.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Cyclic peptides are valuable drug candidates for modulating protein-protein interactions.
• Poor membrane permeability hinders their therapeutic use.
• Computational prediction of permeability can streamline drug development.

Purpose of the Study:

• To benchmark machine learning models for predicting cyclic peptide membrane permeability.
• To evaluate different molecular representations and model architectures.
• To assess prediction performance across regression and classification tasks.

Main Methods:

• Evaluated 13 machine learning models using diverse molecular representations (fingerprints, SMILES, graphs, images).
• Utilized PAMPA permeability data for nearly 6000 cyclic peptides from the CycPeptMPDB database.
• Employed random and scaffold splitting strategies to assess model generalizability.

Main Results:

• Model performance varied significantly based on molecular representation and architecture.
• Graph-based models, especially Directed Message Passing Neural Network (DMPNN), demonstrated superior performance.
• Regression tasks generally yielded better results than classification tasks.
• Scaffold splitting indicated lower generalizability than random splitting.

Conclusions:

• Machine learning, particularly graph-based approaches, shows promise for predicting cyclic peptide permeability.
• Current models offer practical value but further improvements are needed.
• Understanding model generalizability is crucial for reliable predictions.