A curated set and shared test make protein-model building comparable
CryoAtlas filters public structure records into one training set with a fixed benchmark.
A preprint introduces CryoAtlas, a curated dataset and benchmark for building atomic models of protein chains into cryo-EM density maps. It combines and filters public database entries up to March 2026 into 17,809 entries that each pair a map, a model and a sequence. 16,657 form the training split.
The authors also built a benchmark of 618 targets with one evaluation protocol. It measures structural accuracy, how well the model fits the map, and whether the chemistry is valid. They compared recent learning-based methods against two established tools, and looked at how map resolution, sequence length, protein category and sequence similarity affect accuracy.
As a case study, retraining one existing method's network on CryoAtlas improved its benchmark results. The authors call the resource an open, reproducible foundation for future work.
For people who own or use data, the lesson is that curation does the work. The records were already public. Filtering them, fixing the split and agreeing the test is what lets different teams compare results and what makes the set worth reusing.
