A Multi-Modal Cyclic Peptide Representation Learning Framework for Membrane Permeability Prediction

Area of Science:

• Computational chemistry
• Drug discovery
• Molecular modeling

Background:

• Cyclic peptides offer therapeutic potential but face challenges in predicting membrane permeability.
• Current prediction methods using SMILES, graphs, or 3D structures have limitations in capturing comprehensive molecular information.
• Existing multimodal approaches struggle with effective integration of heterogeneous data.

Purpose of the Study:

• To develop an advanced multimodal framework, MultiMol, for accurate prediction of cyclic peptide membrane permeability.
• To overcome limitations of existing methods by integrating diverse molecular representations.
• To enhance the efficiency of virtual screening for drug candidates.

Main Methods:

• MultiMol integrates SMILES sequences, molecular images, molecular graphs, and 3D conformations.
• The framework utilizes tailored pre-training tasks and a scalable fusion mechanism for data integration.
• Performance was evaluated against existing state-of-the-art methods.

Main Results:

• MultiMol demonstrated superior performance in predicting cyclic peptide membrane permeability compared to existing methods.
• Visualization and interpretability analyses confirmed strong feature extraction and generalization capabilities.
• The framework successfully prioritized potential KRAS-targeting cyclic peptides.

Conclusions:

• MultiMol provides a robust and accurate solution for cyclic peptide membrane permeability prediction.
• The framework's ability to integrate multimodal data enhances drug discovery and virtual screening processes.
• MultiMol offers practical utility in identifying promising drug candidates for therapeutic applications.