Historically, researchers discovered new drugs by testing the therapeutic effects of compounds on biological systems before they knew the exact biological target. However, a series of events in the 1990s drastically altered this strategy.
“We started being able to make proteins. We started being able to take pictures of those proteins with X-ray crystallography, and we started being able to model those pictures in a computer,” said chemical biologist Alex Federation, who is also the chief executive officer and cofounder of the therapeutics company Talus Bioscience, Inc. “That’s the time that…the structure-based revolution started.”
Alex Federation is the chief executive officer and cofounder of Talus Biosciences, Inc.
Talus Biosciences, Inc.
As part of this revolution, researchers designed drug discovery tools to model drug-target interactions based on the proteins’ 3D structures. However, Federation noted that this approach of using such co-folding models does not work for about 40 percent of the proteome, for which scientists do not know the protein structures.
Computational biologist Lindsay Pino, the chief technology officer and cofounder of Talus Bioscience, Inc., added, “The protein databank really enabled a lot of the co-folding models [that predict protein structures] to be developed. [But] a lot of the…proteins aren’t well represented in those databases.”
Since this dark proteome consists of disease-relevant proteins, Pino and Federation sought to bridge the gap by developing a structure-free AI model. The platform, called Ptarmigan-1, predicts interactions between all proteins—including unstructured proteins that lack a fixed 3D structure and those conventionally considered undruggable—in the human body, enabling proteome-wide drug discovery.1
Mapping the Unstructured Proteome
Conventionally, researchers isolate proteins from cells and subject them to methods like X-ray crystallography to obtain their 3D structures. However, unstructured proteins, also called intrinsically disordered proteins, function in their native state within cells but don’t retain these shapes well when they are removed from this environment.2
To preserve the cellular context in which these proteins occur, Pino, Federation, and their team turned to human cell lines. By subjecting these cells to a proprietary proteomics technology, the researchers analyzed proteins in their native context to track protein dynamics in response to compounds.
However, the process of obtaining this data was not straightforward. “It’s kind of like a whack-a-mole thing,” remarked Pino. “First, you figure out how to physically prepare more samples…Now you need more mass spectrometers, or you need to get them to run faster. And then, you need to figure out how you are going to store all this data and harmonize all this data together.”
Lindsay Pino is the chief technology officer and cofounder of Talus Biosciences, Inc.
Talus Biosciences, Inc.
The team overcame these challenges using automated proteomic approaches and then used the data to train Ptarmigan-1, which is fundamentally different than most conventional AI-based drug discovery tools. These traditional models compress multidimensional information about a protein, including its structure, into a single numerical representation that the model can process. However, this approach could blur details about individual protein regions.
Conversely, Ptarmigan-1 retains information about individual amino acids to create maps that can predict the likelihood of a ligand interacting with an amino acid residue regardless of its 3D shape. This structure-agnostic approach also makes it faster than other structure-dependent AI models.
“A co-folding model takes about a couple of days to screen a few thousand compounds that might bind to a target,” said Federation. But by bypassing protein folding, Ptarmigan-1 helped the researchers screen the entire proteome against more than a billion compounds in a day.
From AI Predictions to Potential Therapeutics
As a proof of concept, Pino, Federation, and their team investigated the model’s utility in identifying new inhibitors of the transcription factor signal transducer and activator of transcription 6 (STAT6) by comparing it to another co-folding model. The inhibitors for STAT6, which is a validated drug target against inflammatory conditions, were recently announced in a patent, but neither Ptarmigan-1 nor the co-folding model had seen this information during training.
In their study, the team showed that Ptarmigan-1 accurately identified successful drug candidates that the co-folding model could not. The model also outperformed the co-folding model by indicating the drugs’ target sites on STAT6.Encouraged by this success, the researchers allowed other scientists to access to Ptarmigan-1 through a public portal. “In the first two weeks of putting the model out, people ran about two million experiments,” said Federation.
“It’s been really interesting just to see how they’re using it, what questions they’re coming to us with,” said Pino. “There are thousands of these unstructured proteins; [that] is way more than any one lab can work on. That’s also why I’m excited to get it into more people’s hands.”
Both Pino and Federation noted that they will continue finetuning the model such that scientists can take the lead candidates that Ptarmigan-1 suggests toward the clinic. “We’re seeing progress in real time, which is a really great feeling,” said Federation. “We’re excited and deploying it in our labs and other labs as well, but just, seeing that [progress] is pretty awesome.”