A decoupled alignment kernel for peptide membrane permeability predictions

Area of Science:

• Computational chemistry and bioinformatics
• Machine learning for molecular modeling
• Peptide drug discovery

Background:

• Cyclic peptides show promise for targeting intracellular sites, but cell-membrane permeability is a major challenge.
• Limited data and the need for reliable uncertainty quantification hinder progress in this area.
• Existing deep learning models are often data-hungry and complex.

Purpose of the Study:

• To introduce novel, data-efficient kernel methods for modeling cyclic peptide permeability.
• To improve the accuracy and reliability of uncertainty estimation in permeability predictions.
• To provide a robust framework for analyzing sequence-level features of cyclic peptides.

Main Methods:

• Developed monomer-aware decoupled global alignment kernels (MD-GAK) coupling chemical similarity with sequence alignment.
• Introduced a variant, position-aware PMD-GAK, incorporating positional priors for enhanced calibration.
• Utilized Gaussian Processes as the predictive model for uncertainty estimation.

Main Results:

• MD-GAK and PMD-GAK demonstrated superior performance compared to state-of-the-art models across multiple metrics.
• PMD-GAK showed improved calibration accuracy, reducing uncertainty estimation errors.
• The alignment-aware Gaussian Processes provided better discrimination and scaffold-level robustness.

Conclusions:

• MD-GAK and PMD-GAK offer effective, computationally efficient solutions for predicting cyclic peptide permeability.
• These methods enhance sequence-level analysis and probabilistic modeling for peptide design.
• The framework provides a valuable tool for advancing cyclic peptide-based therapeutics.