Penn Engineers Build Open-Source Platform to Predict Peptide Drug Properties

A team at University of Pennsylvania has created PeptiVerse, an AI-powered platform that predicts chemical and biological properties of peptides, the amino acid chains behind therapies such as GLP-1 weight-loss drugs. Described in Nature Communications, the tool was trained on a wide range of data sets so it can estimate traits that determine whether a peptide could work as a drug, including its solubility, ability to enter cells, toxicity, and how long it remains active in the body.

Existing prediction tools tend to cover a narrower set of traits or a single type of peptide. PeptiVerse instead brings many predictions together in one open-source, accessible platform that can evaluate both standard peptides and chemically modified versions engineered to perform better as therapeutics.

According to senior author Pranam Chatterjee, binding to the right target is only part of what makes a peptide useful, and that discovering too late that a promising molecule cannot become a medicine is one of the worst outcomes in drug discovery. PeptiVerse, he said, lets researchers check many of those make-or-break properties earlier, before committing the time and resources needed to synthesize and test candidates.

Building the platform required gathering data scattered across separate studies covering solubility, cell entry, red blood cell damage, protein buildup, and long-term stability. Co-author Sophia Vincoff said each data set came from a different type of experiment, requiring the team to standardize the data and compare multiple model architectures to find the best-performing approach for each property, rather than relying on one model for every task. 

Unlike typical computational tools that require programming skills to run, PeptiVerse includes a web interface where users type in a peptide sequence, select properties, and view predictions through a visual dashboard. Yinuo Zhang, the paper's first author, said the interface lets researchers interact with the platform directly rather than download a Python package. Users can also view the training data behind each prediction to compare their peptides against previously characterized molecules.

Beyond screening, PeptiVerse can pair with generative AI tools that design new peptides, an approach already used in the Chatterjee Lab's PepTune, TR2-D2, MOG-DFM, and moPPIt projects. Chatterjee said the property predictions can guide generative models toward molecules with desired characteristics from the start, rather than evaluating candidates only after they are generated. The open-source platform is designed to keep expanding as more peptide data becomes available and the community contributes new data sets and models.