Penn Engineers Launch PeptiVerse, AI Platform to Accelerate Peptide Drug Discovery

A team of engineers at Penn has built something that could reshape how new peptide drugs get discovered. They call it PeptiVerse, and it's an artificial intelligence platform designed to predict whether a peptide—a string of amino acids—has the chemical and biological properties needed to become a working medicine. The work appears in Nature Communications, and it arrives at a moment when peptides have moved from theoretical promise into real-world success. GLP-1 drugs, the weight-loss treatments now widely prescribed, have proven that peptides can do what we need them to do. But getting from a promising molecule to an actual drug is a long, expensive road, and PeptiVerse is meant to shorten it.

The problem it solves is straightforward but real. Researchers working on peptide drugs need to know whether a candidate molecule will dissolve in the right way, whether it can cross into cells, whether it will poison the body, whether it will stick around long enough to do its job. These are make-or-break questions. Right now, researchers find answers by synthesizing peptides and testing them in the lab—a process that costs time and money. The worst outcome, as Pranam Chatterjee, the senior author and an assistant professor in bioengineering and computer science at Penn, puts it, is discovering too late that a molecule that looked promising cannot actually become a medicine. PeptiVerse lets researchers check these critical properties early, before they commit resources to synthesis and testing.

What makes PeptiVerse different from existing tools is its scope and accessibility. Other platforms exist that predict peptide properties, but they tend to focus narrowly—one tool for solubility, another for cell entry, another for toxicity. PeptiVerse brings many of those predictions together in a single, open-source platform. More than that, it includes a web interface. You type in a peptide sequence, select the properties you want to know about, and the system shows you the predictions on a visual dashboard. This matters because many researchers working with peptides are biologists, not computer scientists. They need answers, not code to download and debug.

Building the platform required work that rarely gets noticed. The researchers had to hunt down data scattered across separate studies and experiments. Some data sets measured solubility. Others measured cell permeability, toxicity to red blood cells, resistance to unwanted protein buildup, or how long peptides remain active. Each data set came from a different type of experiment, with different methods and standards. Sophia Vincoff, a doctoral student who worked on the project, explains that the team had to understand what each experiment was actually measuring, standardize the data, and organize it so machine-learning models could learn from it fairly. They tested many different model architectures and found that the strongest performer varied by task. Sometimes simpler models worked as well as or better than sophisticated ones. The result is a toolkit where each prediction uses the model that works best for that specific job.

The platform is already being used in two ways. First, as a screening tool. Instead of synthesizing and testing peptide candidates one by one, researchers can now use PeptiVerse to quickly assess which molecules look most promising before moving into the lab. Second, and more ambitiously, it can work alongside generative AI tools that propose entirely new peptides. In this mode, PeptiVerse's property predictions guide the generative models toward molecules with desired characteristics—stronger binding, fewer unwanted interactions with other proteins, better solubility, greater ability to cross cell membranes, lower toxicity. The predictions become part of the search itself, not just a filter applied afterward. Chatterjee's lab has already used this approach to develop new peptide-design innovations with names like PepTune, TR2-D2, MOG-DFM, and moPPIt.

What makes PeptiVerse genuinely open is that it was designed to grow. As more peptide data becomes available, the platform can be updated with new data sets, improved models, and additional properties. Some peptide properties are still harder to predict than others, largely because there is less experimental data available. More measurements could help PeptiVerse improve existing predictions and add new ones—for instance, whether a peptide is likely to activate or block a receptor. Because the platform is open-source, academic labs can use it, contribute data, or build new predictors. Companies developing peptide therapeutics could use the framework with their own internal data, creating customized versions tailored to the molecules they are trying to develop. Chatterjee notes that the universe of possible peptides is too large for any one lab to map alone. PeptiVerse was built so that other researchers can help expand the map, adding new data and models that accelerate the search for new and better peptide drugs.