A protein’s sequence does not, by itself, tell us the three-dimensional shape it will form. Structure prediction proposes that shape computationally. If predictions are selected or their coordinates changed after the experimental answer appears, it becomes difficult to establish what was predicted beforehand. I prepared a fixed package of prediction files and validation records before comparison with the experimental structure.
The freeze concerns files, not a physical protein. Predicted atomic positions are stored as coordinates, and the record identifies the files intended for comparison. That makes it possible to return to the prepared prediction rather than choose whichever file later receives the best score.
Checking the files before comparing the structures
A coordinate file still needs to be checked. I preserved a validation package that checks whether its residues match the input protein sequence, whether chains are distinguished, and whether the coordinates follow the submission format. A residue is an amino-acid unit within a protein chain. Errors in these labels can lead to comparisons between different objects even when the filenames suggest the same target.
The package also records a SHA-256 hash for each file. This long identifier is calculated from the file’s contents and helps check whether a later copy matches the preserved file. It can distinguish files with the same name but different contents. It does not measure structural accuracy: checking file identity and format leaves the comparison with an experimental structure still to be done.
Five models submitted to CASP17
On 20 August 2026, CASP17, a protein-structure prediction challenge, received five T2461 models I submitted as YooIndependent. The work had moved beyond file preparation to official submission. On 20 September, I also submitted a methods abstract describing template-guided assembly and interface repacking of an ESMFold model.
The submissions are listed in the official CASP17 prediction record. As of 20 September, I have not confirmed an official assessment score or ranking; the number of models received does not establish prediction accuracy. Candidate coordinates and unpublished numerical results remain private. When assessment results are available, I plan to compare them against the preserved submission files.