The AlphaFold team described a neural-network system that predicts protein structures from amino-acid sequences with accuracy that was competitive with experimental structures in the CASP14 assessment.
Structure predictions can help researchers form hypotheses about function, interactions and mutations, and can prioritize experiments. They do not replace laboratory measurements, especially for flexible proteins, complexes or context-dependent states.
The broader lesson is that computation can expand the reachable search space while preserving the need for validation. A predicted structure is evidence for a question, not a clinical conclusion.
STKEY note: editorial summary of the research publication, not a reproduction of its figures or database.