PCPpred: A Large Language Model-Based Tool for Predicting Cyclopeptide Membrane Permeability

PCPpred is an open-source tool developed by the Raghava research group at the Indian Institute of Information Technology Delhi (IIIT Delhi), specifically designed for cyclopeptide drug development. It uses large language models and ensemble learning techniques to predict the membrane permeability of chemically modified peptides. This tool supports predictions for four mainstream permeability experimental models—PAMPA, Caco-2, RRCK, and MDCK—and provides sequence conversion functionality from MAP format to SMILES/HELM. Its aim is to lower the barrier to oral cyclopeptide drug design and accelerate the development of related therapeutic modalities.

Peptide drugs have great potential due to their high specificity and low toxicity, but oral administration is a bottleneck in clinical applications—their large molecular weight and strong polarity lead to extremely low bioavailability. Traditional linear peptides are easily degraded; cyclopeptides improve metabolic stability through cyclization, but cyclization does not guarantee membrane permeability, so chemical modifications (such as N-methylation and introduction of non-natural residues) are key. Experimental determination of membrane permeability is costly and time-consuming, so developing accurate prediction tools is of great value.

Multidimensional Molecular Representation System

PCPpred integrates four types of molecular representations:

1. Molecular descriptors (physicochemical properties such as molecular weight and lipid-water partition coefficient)
2. Molecular fingerprints (e.g., Klekota-Roth fingerprints to capture substructures)
3. Molecular embeddings (pre-trained language models learn distributed representations from SMILES)
4. Atom-level features (fine-grained information like atom type and hybridization state)

Ensemble Learning Prediction Architecture

It adopts a stacked ensemble strategy, combining multiple base learners (such as LightGBM, XGBoost, random forest, etc.), and optimizes weights through a meta-learner to reduce overfitting and improve stability.

Supported Experimental Models

Covers four mainstream in vitro models:

• PAMPA: Simulates passive transmembrane diffusion
• Caco-2: Simulates the small intestinal epithelial barrier (FDA-approved)
• RRCK: A simplified version of Caco-2 for rapid screening
• MDCK: Canine kidney cell model for supplementary validation

MAP format is commonly used for annotating cyclopeptide modifications. PCPpred provides two scripts for processing:

1. map_to_smiles.py: Converts MAP sequences to universal SMILES strings
2. map_to_helm.py: Batch converts to HELM format (a biopharmaceutical industry standard) for easy integration with commercial software

Example input: {nnr:ABU}{nnr:0OZ}{nnr:9XD}V{nnr:9XD}AA{d}{nnr:9XD}{nnr:9XD}{nnr:0Q3}{nnr:MBM}{cyc:N-C}

Oral Peptide Drug Design

• High-throughput screening of virtual compound libraries
• Evaluation of the impact of chemical modifications on permeability
• Identification of cyclopeptide scaffolds with oral potential
• Guidance on synthesis priority ranking

Structure-Permeability Relationship Research

• Analysis of the relationship between ring size, amino acid residues, hydrophobicity, etc., and permeability

Peptide Drug Repurposing

Assists in evaluating the feasibility of converting injectable peptides to oral forms and the required modification strategies

Limitations

1. Limited coverage of training data; predictions for extremely novel modifications may have deviations
2. There is a gap between in vitro models and human bioavailability
3. Limited consideration of active transport (e.g., P-glycoprotein)

Future Directions

• Expand the training dataset
• Introduce physicochemical simulations to enhance mechanistic understanding
• Develop models that consider transporter protein interactions
• Establish an end-to-end prediction process for oral bioavailability

PCPpred combines computational chemistry and large language model technology, providing an open-source, customizable permeability prediction tool. It lowers the barrier to oral cyclopeptide drug design and is expected to accelerate the development of this field, making it a computational resource worth attention for practitioners in academia and industry.