Penn Engineers have developed PeptiVerse , an AI-powered platform that predicts key chemical and biological properties of peptides, the strings of amino acids whose medical potential has been demonstrated by the success of GLP-1 drugs, the widely used weight-loss treatments.
In Nature Communications , the researchers describe how they trained PeptiVerse using a wide range of data sets, allowing it to predict properties that can help determine whether a peptide is worth pursuing as a potential drug, such as the peptide’s likelihood of dissolving, entering cells, avoiding toxicity and lasting long enough in the body to have an effect.
While tools for predicting such properties exist, those tools often focus on a narrower set of traits or only one kind of peptide. PeptiVerse, by contrast, brings many of those predictions together in one open-source, easily accessible platform, allowing users to evaluate both ordinary peptides and chemically modified versions designed to work better as drugs.
“Peptide drugs have enormous potential, but binding to the right target is only one part of what makes a molecule useful,” says Pranam Chatterjee , Africk-Lesley Distinguished Scholar of Innovation in Engineering, Assistant Professor in Bioengineering (BE) and in Computer and Information Science (CIS), and senior author of the new study.
“In drug discovery, one of the worst outcomes is finding out too late that a promising molecule cannot actually become a medicine,” says Chatterjee. “PeptiVerse gives researchers a way to check many of those make-or-break properties earlier, before they invest the time and resources required to synthesize and test candidate drugs.”
Putting Peptide Predictions in One Place
To build PeptiVerse, the researchers first had to gather data that had been scattered across separate studies. Some data sets described whether peptides dissolve easily. Others measured whether peptides enter cells, avoid damaging red blood cells, resist unwanted protein buildup or remain active over time.
Synthesizing the data sets required far more than saving them in the same location. “Each data set came from a different type of experiment,” says Sophia Vincoff , a doctoral student in BE and co-author of the study. “We had to understand what each experiment was measuring, standardize the data for each property and organize it so machine-learning models could learn from it and be tested fairly. Then we compared many model architectures to find the best approach for each prediction task.”
Rather than assuming that one model works best for every task, PeptiVerse uses the strongest-performing model for each property. In some cases, simpler models performed as well as or better than more sophisticated approaches. In others, the opposite was true.
“We wanted PeptiVerse to function like a toolkit,” says Yinuo Zhang , a postdoctoral researcher in BE and the paper’s first author. “Instead of having each predictor live separately, PeptiVerse puts them together in a platform that can grow as new data and models become available.”
Making Peptide Predictions Easier to Use
Most new computational tools come in the form of libraries or packages, bundles of code that researchers can download and run on their own computers, but that typically require some level of facility with programming to operate.
PeptiVerse, by contrast, includes a web interface that allows users to simply type in a peptide sequence, select properties and receive predictions through a visual dashboard.
That design makes the platform especially useful for experimental researchers, like biologists, who may have peptide candidates they are already studying and want a fast way to assess whether those molecules are likely to have the properties needed for further testing.
“I come from a biology perspective, and I’m a very visual person,” says Zhang. “We wanted researchers to be able to see what was going on, not just download a Python package. The interface makes PeptiVerse something people can interact with directly.”
The platform also lets users view the data used to train the models, helping them understand how their own peptide candidates compare with molecules that have already been experimentally characterized.
“For experimentalists, it matters where the data came from,” says Vincoff. “PeptiVerse makes it easier to look at the data behind the predictions, which helps users interpret what the models are telling them.”
From Screening to Generating Drug Candidates
One immediate use of PeptiVerse is as a screening tool. Rather than synthesizing and testing peptide candidates one by one, researchers in academia and industry alike can use the platform to quickly assess which molecules appear most promising before moving into the lab.
The platform could also play a more ambitious role in AI-driven drug discovery. In addition to evaluating existing candidates, PeptiVerse can be paired with generative AI tools that propose entirely new peptides. That approach has already helped the Chatterjee Lab develop new peptide-design innovations, including PepTune , TR2-D2 , MOG-DFM and moPPIt , which use property predictions from PeptiVerse to guide the design of peptides with desired therapeutic features.
In this paradigm, rather than serving only as a filter to screen existing peptides, PeptiVerse can shape the search for new drug candidates. The system’s powerful predictions allow generative AI models to prioritize candidates with desirable characteristics — such as stronger binding, reduced nonspecific interactions with other proteins, better solubility, greater permeability and lower toxicity — before selecting particular molecules for synthesis and real-world testing.
“The property predictions become part of the search itself,” says Chatterjee. “Instead of evaluating peptides only after they are generated, we can use those predictions to guide generative AI models toward molecules with the characteristics we want from the very beginning.”
An Expanding Universe of Peptide Data
The researchers also designed PeptiVerse to keep growing. As more peptide data becomes available, the platform can be updated with new data sets, improved models and additional properties. That matters because some peptide properties are still harder to predict than others, largely because there is less experimental data available to train models about them.
More measurements could help PeptiVerse improve existing predictions and add new ones, such as whether a peptide is likely to activate or block a receptor, activating or shutting down a particular biochemical pathway. “PeptiVerse is not meant to be finished,” says Chatterjee. “The more data the community contributes, the better these models can become.”
Since PeptiVerse is open source, researchers can adapt it for their own purposes. Academic labs can use the existing platform, contribute data or build new predictors. Companies developing peptide therapeutics could also use the framework with their own internal data sets, creating versions of PeptiVerse tailored to the kinds of molecules they are trying to develop.
“The universe of possible peptides is too large for any one lab to map on its own,” says Chatterjee. “We built PeptiVerse so that other researchers can help expand the map, adding new data and models that accelerate the search for new and better peptide drugs.”
This study was conducted at the University of Pennsylvania School of Engineering and Applied Science (Penn Engineering) and supported by the Hartwell Foundation, the National Institutes of Health (R35GM155282) and by a pilot grant from the High-throughput Institute for Discovery (HIT-ID) at the University of Pennsylvania.
Additional co-authors include Sophia Tang , Tong Chen and Elizabeth Mahood of Penn Engineering.
Nature Communications
Computational simulation/modeling
Not applicable
PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction
21-Jul-2026
Pranam Chatterjee is a co-founder of Gameto, Inc., UbiquiTx, Inc., AtomBioworks, Inc., and Recognition Bio., Inc., and advises companies involved in peptide therapeutics development. Chatterjee's interests are reviewed and managed by the University of Pennsylvania in accordance with their conflict-of-interest policies. The remaining authors have no conflicts of interest to declare.