Predict a protein structure

Check first whether the answer already exists. If your protein is a known one, the AlphaFold database very likely has a model already, and an experimental structure in the PDB beats any prediction. Predicting something that has been crystallised is a common waste of a week.

For a single chain or a straightforward complex, AlphaFold Server is the path of least resistance: no installation, no queue worth mentioning, and the current model. Its limits are the ones that matter for real work — a daily job cap, a ceiling on total residues, and the fact that your sequence goes to someone else’s machine. For an unpublished or commercially sensitive sequence, that last point may decide it on its own.

ColabFold is the answer when the server’s limits bite. It runs the same class of model with the parameters exposed, so you can raise the recycle count on a difficult target, feed a custom multiple sequence alignment, or batch through a hundred sequences unattended. The cost is that you are managing a notebook and a runtime.

The older threading servers such as I-TASSER have not been superseded for every case. They remain useful when a good template exists and you want a model built explicitly on it, and their output includes functional inference — ligand and binding site predictions — that the folding models do not attempt.