AI-driven interpretability enables permeability-aware peptide design at scale

Cyclic Peptide Permeability Modeling

Cyclic peptides have emerged as a promising therapeutic modality capable of targeting challenging intracellular proteins and protein–protein interactions that are often inaccessible to conventional small molecules and biologics. However, despite their favorable potency and selectivity profiles, limited membrane permeability remains one of the most significant barriers to their successful development. A cyclic peptide’s ability to cross biological membranes directly influences oral bioavailability, intracellular target engagement, and overall therapeutic efficacy, making permeability a critical developability parameter in modern peptide drug discovery.

Unlike traditional drug molecules, cyclic peptides exist as dynamic conformational ensembles that continuously adapt to their surrounding environment. Their permeability is governed not only by chemical composition but also by solvent-dependent conformational changes, intramolecular hydrogen bonding, and the ability to transition between polar and nonpolar environments during membrane transport. Experimental permeability assessment across diverse solvent and membrane-mimetic conditions is resource-intensive and difficult to scale, creating a major bottleneck for peptide discovery programs. As pharmaceutical companies increasingly pursue orally available and cell-permeable cyclic peptide therapeutics, there is a growing need for high-throughput computational approaches that can accurately predict permeability by capturing the complex interplay between conformation, solvent environment, and membrane interactions.

Overview

Cyclic peptide membrane permeability is a key determinant of their oral bioavailability and therapeutic potential, critical to predict and optimize early in drug discovery. This case study highlights an interpretable AI-driven workflow for predicting membrane permeability of cyclic peptides across diverse chemical spaces. It was validated on a large-scale PAMPA dataset. The model combines graph-based prediction with an interactive UI that interprets permeability drivers and generates counterfactual modification suggestions for low-permeability peptides, enabling early-stage triage, rational redesign, and confident prioritization of permeable candidates.

Design Challenge

Membrane permeability remains a key bottleneck in cyclic peptide drug discovery due to complex structure–property relationships and limited interpretability of predictive models, making it difficult to identify liabilities early and guide rational optimization strategies.

Objectives

• – Predict membrane permeability of cyclic peptides using large-scale experimental data
• – Enable early identification of permeability liabilities
• – Provide interpretable insights into key drivers of permeability
• – Support rational redesign through counterfactual modification suggestions
• – Prioritize permeable candidates for downstream development

Outcomes

Effective Permeability Prediction Performance

~99.6% sensitivity with strong enrichment of permeable candidates

• 68% overall classification accuracy on test data
• ~99.6% sensitivity for identifying permeable peptides
• Strong enrichment of high-permeability candidates for prioritization
• Actionable insights to guide redesign of impermeable leads

This workflow enables early-stage permeability triage, candidate prioritization, and data-driven optimization for cyclic peptide discovery.

More on Aganitha’s Peptide Engineering Pipeline: Click here.