Talus Bio, a Seattle-based biotech company, claims its new drug screening model can accelerate the discovery of treatments for previously “undruggable” proteins by skipping the costly and time-consuming step of protein folding.
Targeting Undruggable Proteins
Most drug discovery relies on accurately predicting a protein’s 3D structure, a task made famous by AlphaFold, which has mapped over 200 million proteins. However, many proteins, particularly transcription factors that regulate gene activity, lack stable structures for modeling. These flexible proteins shift shape, leaving no fixed binding pocket for drugs to target.
Alex Federation, CEO of Talus Bio, stated that scientists and oncologists often express a desire for drugs targeting transcription factors, which regulate gene activity. While these proteins are linked to cancer and other diseases, pharmaceutical companies have long deemed them undruggable.
Read Also: Expanded Pharma Tariffs Target Smaller Drugmakers Over R&D Fears
According to Lindsay Pino, Talus’s chief technology officer, roughly half of human proteins lack a stable structure suitable for folding models. This gap helps explain why, of an estimated 20,000 human proteins, 87% still lack an approved drug or potent small-molecule ligand, per a July preprint from Talus.
A Plate of Spaghetti Approach
Rather than attempting to predict protein structures, Talus’s model, Ptarmigan-1, learns from mass spectrometry data showing where compounds bind proteins inside cells. “If you try to fold a transcription factor, you just get a plate of spaghetti,” Pino said. “You can’t do drug discovery on a plate of spaghetti.”
Talus’s model also avoids the escalating costs of traditional screening. Federation noted that co-folding a protein with potential drugs is affordable for a single compound but becomes prohibitively expensive at scale. A typical drug discovery library contains about a million compounds, which could cost $100,000 to $1 million and take months to screen for a single protein using folding-based methods.
Ptarmigan-1 processes each compound in roughly 10 milliseconds, compared to 54 seconds for Boltz-2, an open-source model that predicts molecular structures and binding affinity, according to Talus’s preprint. This 5,000-fold speed advantage could reduce a million-compound screen from nearly two years to under three hours.
Read Also: FDA finalizes guidance on written responses and meeting types
Benchmarks and Real-World Testing
In a retrospective test on STAT6, a transcription factor, Ptarmigan-1 identified 40 inhibitors from Pfizer patents with an area under the curve (AUC) of 0.94. Boltz-2 and a standard docking method both scored 0.58, near chance levels. On well-folded targets, Boltz-2 outperformed Ptarmigan-1, but Talus positions the two models as complementary: Ptarmigan-1 quickly narrows billions of compounds, while structural models refine the top candidates.
The company’s compound library enabled it to retrieve the highest predicted binders for all 20,431 human proteins from 3.4 billion compounds in under a day using 20 Nvidia H100 GPU-hours.
“We’ve been sitting on these targets we’ve known about for decades, that we know are good targets,” Pino said. “We just haven’t had the tools to find them yet.”
