Machine Learning Tool Predicts Which Peptides Can Penetrate Cells for Drug Delivery | Research

The takeaway

PerseuCPP, a new machine learning tool, predicts cell-penetrating peptides with 98.4% accuracy (AUC) and outperforms existing methods, accelerating peptide-based drug delivery design.

AUC 0.984
PerseuCPP achieved near-perfect discrimination between cell-penetrating and non-penetrating peptides, surpassing all existing prediction methods and providing a reliable computational screening tool.

What the researchers found

The CPP prediction model achieved superior performance compared to existing state-of-the-art methods: MCC of 0.854, Recall of 0.860, and AUC of 0.984. The model was trained on a balanced dataset of 967 CPPs and non-CPPs with 10-fold cross-validation and validated on two independent test sets.

The uptake efficiency predictor — trained on 140 CPPs — achieved competitive results with Recall of 0.761 and AUC of 0.690. The model is interpretable, identifying which physicochemical properties, structural features, and atomic compositions are most important for cell penetration. The tool uses Extremely Randomized Trees as its core algorithm.

Why it matters

Drug delivery is one of the biggest challenges in medicine — many promising drugs can't reach their targets inside cells. Cell-penetrating peptides solve this problem but discovering new ones through lab experiments is expensive and slow. PerseuCPP dramatically accelerates this discovery process by computationally screening peptide candidates before any lab work, potentially reducing costs and time in peptide drug delivery development. Its interpretability also teaches researchers what makes a peptide good at entering cells.

How the study worked

The researchers built a two-stage machine learning pipeline. Stage one identifies CPPs using an Extremely Randomized Trees classifier trained on 967 peptides (balanced CPPs and non-CPPs) with descriptors capturing physicochemical properties, structural features, and atomic composition. Ten-fold cross-validation was used for training, with two independent datasets for validation. Stage two predicts uptake efficiency using a similar approach trained on 140 peptides. Both stages were benchmarked against existing prediction methods.

What this study cannot tell us

The uptake efficiency predictor had notably lower performance (AUC 0.690) compared to the CPP identification stage (AUC 0.984), likely due to the small training set of only 140 peptides. Computational predictions need experimental validation — a predicted CPP may not work as expected in actual biological systems. The model was trained on known CPP datasets that may not represent the full diversity of potential cell-penetrating peptides. Performance on novel, structurally unusual peptides is untested.

How to read the evidence

This is a computational methods paper demonstrating a machine learning tool with rigorous validation — 10-fold cross-validation plus independent test sets. The CPP prediction performance is strong, though the efficiency predictor is weaker. As a bioinformatics tool rather than a clinical study, evidence quality is assessed by prediction accuracy and generalizability rather than clinical outcomes.

When this study was published

Published in 2025, this study uses current machine learning approaches and up-to-date CPP datasets. The tool is freely available on GitHub for immediate use by the research community.

The bigger picture

Cell-penetrating peptides are becoming increasingly important as drug delivery vehicles, particularly for challenging therapeutic cargoes like nucleic acids, proteins, and nanoparticles that can't cross cell membranes on their own. Machine learning is transforming peptide design across the field — from antimicrobial peptide prediction to protein structure forecasting. PerseuCPP joins a growing toolkit of computational methods that make peptide-based drug development faster, cheaper, and more rational.

Questions still open

Can PerseuCPP's efficiency predictions be improved with larger training datasets as more CPP uptake data becomes available?
How well does the model perform on cyclic peptides, stapled peptides, and other non-linear peptide architectures?
Could PerseuCPP be integrated with other computational tools to design optimal CPPs for specific therapeutic cargoes?

Common questions

What are cell-penetrating peptides and why are they important for medicine?

Cell-penetrating peptides (CPPs) are short protein fragments that can pass through cell membranes without damaging them. This makes them valuable delivery vehicles — they can carry drugs, genes, or proteins directly into cells that would otherwise be unreachable. This is especially important for therapies like gene therapy and RNA-based drugs that need to get inside cells to work.

How does PerseuCPP predict whether a peptide can enter cells?

PerseuCPP analyzes the chemical and physical properties of a peptide — things like its charge, size, shape, and atomic makeup — and uses machine learning to compare these features against thousands of known cell-penetrating and non-penetrating peptides. It then predicts both whether the peptide can enter cells and how efficiently it does so. The tool also explains which features are most important, helping researchers design better CPPs.

Read the original research

PerseuCPP: a machine learning strategy to predict cell-penetrating peptides and their uptake efficiency.

Bioinformatics advances, 5(1), vbaf213

View on PubMed ↗Publisher’s paper ↗

Citation

Bernardes-Loch, Rayane Monique; de Oliveira Almeida, Gustavo; Brasiliano, Igor Teixeira; Meira, Wagner; Pires, Douglas E V; Baracat-Pereira, Maria Cristina; de Azevedo Silveira, Sabrina. (2025). PerseuCPP: a machine learning strategy to predict cell-penetrating peptides and their uptake efficiency.. Bioinformatics advances, 5(1), vbaf213. https://doi.org/10.1093/bioadv/vbaf213