AI Platform Maps Drug Interactions Across the Entire Human Proteome

Historically, researchers discovered new drugs by testing the therapeutic effects of compounds on biological systems before identifying the exact biological target. A series of developments in the 1990s fundamentally changed 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 Alex Federation, chemical biologist and chief executive officer and cofounder of Talus Bioscience, Inc. “That is when the structure-based revolution began.”

This structure-based revolution produced drug discovery tools that model drug-target interactions using proteins’ three-dimensional structures. However, Federation noted that co-folding models do not work for roughly 40 percent of the proteome, because scientists do not know the structures for those proteins.

Lindsay Pino, computational biologist and chief technology officer and cofounder of Talus Bioscience, Inc., added, “The protein databank enabled many of the co-folding models that predict protein structures to be developed. Yet many proteins are not well represented in those databases.”

Because this dark proteome includes disease-relevant proteins, Pino and Federation sought to close the gap by developing a structure-free AI model. The platform, Ptarmigan-1, predicts interactions between all proteins in the human body—including unstructured proteins without a fixed three-dimensional shape and targets conventionally considered undruggable—enabling proteome-wide drug discovery.

Charting the Unstructured Proteome

Conventionally, researchers isolate proteins from cells and use methods such as X-ray crystallography to determine their three-dimensional structures. However, unstructured proteins, also called intrinsically disordered proteins, function in their native cellular state but do not retain those shapes well once removed from that environment.

To preserve the cellular context in which these proteins occur, Pino, Federation, and their team turned to human cell lines. Using a proprietary proteomics technology, the researchers analyzed proteins in their native setting to track protein dynamics in response to compounds.

Generating this data proved challenging. “It’s kind of like a whack-a-mole thing,” Pino remarked. “First, you figure out how to physically prepare more samples…Now you need more mass spectrometers, or you need them to run faster. Then you need to determine how to store all this data and harmonize it.”

The team overcame these obstacles through automated proteomic approaches and used the data to train Ptarmigan-1, which differs fundamentally from most conventional AI drug discovery tools. Traditional models compress multidimensional information about a protein, including its structure, into a single numerical representation for processing. This approach can obscure details about individual protein regions.

By contrast, Ptarmigan-1 retains information about individual amino acids to create maps predicting the likelihood of a ligand interacting with an amino acid residue, regardless of its three-dimensional shape. This structure-agnostic approach also makes it faster than 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,” Federation said. By bypassing protein folding, Ptarmigan-1 enabled researchers to screen the entire proteome against more than a billion compounds in a day.

Translating AI Predictions into Therapeutics

As a proof of concept, Pino, Federation, and their team evaluated the model’s ability to identify new inhibitors of the transcription factor signal transducer and activator of transcription 6 (STAT6) by comparing it to a co-folding model. STAT6 inhibitors, a validated drug target for inflammatory conditions, were recently announced in a patent, but neither Ptarmigan-1 nor the co-folding model had encountered that information during training.

In their study, the team demonstrated that Ptarmigan-1 accurately identified successful drug candidates that the co-folding model missed. The model also outperformed the co-folding model by indicating the drugs’ target sites on STAT6. Encouraged by this success, the researchers provided other scientists access to Ptarmigan-1 through a public portal. “In the first two weeks of making the model available, people ran about two million experiments,” Federation said.

“It’s been interesting to see how they are using it and what questions they bring to us,” Pino said. “There are thousands of these unstructured proteins—far more than any one lab can handle. That is also why I am excited to put it into more people’s hands.”

Both Pino and Federation said they will continue refining the model so scientists can advance Ptarmigan-1’s lead candidates toward clinical application. “We’re seeing progress in real time, which is a great feeling,” Federation said. “We’re excited to deploy it in our labs and other labs, and just seeing that progress is pretty awesome.”

Source link