How-tos for research tools and news about Nadoo Bio.
A flash-attn check on Blackwell pointed to the wrong suspect: the PyTorch 2.7 fp32 attention result we used as a reference. Here are the conditions, version comparisons, production image audit, and a script to check your own environment.
We ran Boltz-2 on 16 public TYK2 candidates and compared its predictions with historical experiments. From two overlapping top-three candidates to a large ranking discrepancy, this case explains molecules, binding, and prediction metrics.
Why does drug discovery need computer science? From choosing experimental candidates to validating research tools, biology, data, and computation shape the same decisions.