No Need to Resolve Protein Structures First: Talus Opens AI Model Ptarmigan-1 for Researchers to Try

The search for drugs often begins with identifying pockets on a protein’s surface that can accommodate small molecules. However, proteins are constantly changing within cells, and some regions are difficult to describe with a fixed three-dimensional structure. On October 1, Talus Bio opened the public Ptarmigan-1 platform, allowing researchers to try predicting small-molecule binding sites directly from protein sequences and seek experimental leads for these difficult-to-characterize targets.

Ptarmigan-1 was announced in late July; the new development is that researchers can now use it themselves. Users can select a human protein and pair it with the platform’s compound library or molecules they upload themselves to obtain rankings of candidate molecules and amino acid positions where binding may occur. Currently, each screen is limited to ten thousand compounds, with a limit of ten screens per day; larger-scale screening requires discussing a collaboration directly with the company.

Its approach converts amino acid residues in proteins and compounds into representations in the same 256-dimensional mathematical space, then learns which molecules may interact with which positions. According to Talus, the underlying experimental data come from its proprietary MARMOT proteomics platform and include information on compound–protein interactions within living human cells. Although the interface can display predicted sites on AlphaFold structures, three-dimensional structures are not required inputs for the model’s predictions.

One of the specific examples the company currently provides is STAT6, which is associated with inflammatory diseases. A technical introduction in July described a test in which 40 active molecules from two Pfizer patents were mixed with decoy compounds with similar properties, used for comparison, to examine whether the model could rank the active molecules near the top. Talus says these active molecules were absent from the training data for both Ptarmigan-1 and Boltz-2; Ptarmigan-1 achieved an area under the ROC curve (AUC) of 0.94, compared with approximately 0.58 for molecular docking methods and Boltz-2. This metric reflects discrimination in a specific test and cannot be interpreted as a success rate for drug candidates. The model also pointed to a shallow groove in STAT6’s SH2 domain as a possible binding site.

A company press release published by BioSpace in October separately claimed that the model had identified new STAT6-binding compounds that were validated by a third-party laboratory. However, the release did not name the laboratory or provide sufficient experimental detail to fully assess the strength of the validation. This result should be considered separately from the retrospective test of the patent molecules described above; republication of the press release also does not mean that another research team has independently reproduced the model’s performance.

Computational speed is another advantage Talus emphasizes. The company claims that the model can screen 3.4 billion compounds against human proteins in one day, at 5,000 times the speed of structure-based methods. These figures are performance claims released by the company and do not represent the usage allowance currently available on the public platform. For drug development, the more practical question is whether rapidly generated candidate lists can improve the chances of finding effective molecules in subsequent experiments.

Ptarmigan-1 was trained using binary labels indicating the presence or absence of interactions, rather than measurements of binding affinity. Predicting that a molecule may bind at a particular site therefore still cannot directly answer how strongly it binds, whether it changes protein function, or whether it can safely become a drug. The public platform gives more researchers an opportunity to test their own targets and chemical series; prospective experiments across targets, independent replication, and peer review are needed next to determine how far this route that does not rely on three-dimensional structures can go.