A Multi-Modal Contrastive Learning Framework for Cyclic Peptide Permeability Prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Biotechnology

Background:

• Cyclic peptides are a growing class of therapeutics.
• Predicting cell membrane permeability is crucial for drug efficacy but challenging.
• Current computational methods struggle with diverse structural information in cyclic peptides.

Purpose of the Study:

• To develop an accurate computational framework for predicting cyclic peptide cell membrane permeability.
• To integrate 1D, 2D, and 3D molecular information using deep learning.
• To accelerate the rational design of cell-permeable cyclic peptide drugs.

Main Methods:

• Introduced MCPerm, a multi-modal deep learning framework.
• Integrated 1D SMILES, 2D topological, and 3D geometric data.
• Utilized a novel modality share and contrastive learning strategy.
• Fine-tuned a pretrained peptide language model and used a graph transformer.
• Employed dual contrastive learning for representational consistency.

Main Results:

• MCPerm achieved state-of-the-art performance on the PAMPA dataset.
• Significantly outperformed existing leading methods in permeability prediction.
• Demonstrated robustness and transferability across Caco-2, MDCK, and RRCK assays.
• Attention-based visualization revealed the model learned key chemical principles.

Conclusions:

• MCPerm provides a robust in silico framework for predicting cyclic peptide permeability.
• The model accelerates the design and discovery of effective peptide therapeutics.
• MCPerm offers insights into the chemical basis of peptide permeability, moving beyond a black-box approach.