Jacob Berlin, PhD, CEO of Terray Therapeutics, explains that the critical element to “get right” for a preclinical drug discovery campaign is identifying molecules that rank best among the candidate set. The AI drug discovery company is building an end-to-end pipeline to bring novel small molecule drugs from unseen chemical space to the clinic. Designing these therapeutics from scratch, without patent information or public datasets, will require evaluating thousands to millions of molecules across the discovery process, from hit identification to lead optimization. To support this goal, Terray has unveiled TerraBind, a small molecule potency prediction model that demonstrates an approximately 20% increase in accuracy and 26 times more efficiency gain when compared against Boltz-2, the widely adopted open-source binding affinity model developed by MIT researchers, who recently turned their models into a business. Correspondingly, TerraBind permits screening 26 times more chemical space using the same resources. The model’s efficiency advantages stack up across the development pipeline, from expanding the therapeutic search space to achieving higher quality candidate selection for synthesis. The work is posted as a preprint on arXiv that has not yet been peer reviewed. Unlike co-folding models, such as Nobel Prize-winning AlphaFold and the Boltz series of models, which use a computationally intensive all-atom approach to determine the structure of a molecule when bound to the target, TerraBind’s architecture bypasses this step to predict drug potency directly. Berlin says there is a disconnect between the pictorial representation of a molecule and therapeutic function. This “snapshot in time” on its own does not provide information on binding affinity nor duration,…
