Sinewgrid Protein trains protein and small-molecule design models.

The number of possible proteins and drug-like molecules is vastly larger than anything a lab can make and test, so progress depends on proposing a small number of good candidates and learning quickly from each result. Sinewgrid Protein is the domain where we would build models that generate designs for a target, such as a protein, a binding site or a property, and rank them with structure-prediction and binding models, with an estimate of how uncertain each ranking is. The compute demand has three parts: a large first pretraining run on sequence and structure collections that needs fast interconnect, many scoring and fine-tuning passes after every lab batch, and inference across thousands of candidate designs for each target. Together they make demand uneven and tied to the lab’s calendar.

A program runs as a loop. We define the target and the assay that will judge it, the model generates and ranks designs, a partner lab makes and tests a small, informative batch, and the measured results update the model before the next batch is chosen. We would buy steady capacity for pretraining and on-demand capacity for the scoring that follows each batch. We report three measures that are fixed before the first run: the hit rate against a baseline the partner already trusts, the number of rounds needed to reach a hit, and the compute spent per round. Biosecurity comes first. Design outputs are screened against risk lists, access is tiered, and capability evaluations are reviewed by outside experts before any release. This domain is at the exploring stage.