AlphaFold made protein shapes predictable — and turned a bottleneck into a starting line

What the study found

In 2021, a team reported AlphaFold, a deep-learning system that predicts a protein's three-dimensional structure directly from its amino-acid sequence. In the CASP14 community assessment, its predictions reached accuracy competitive with experimental methods for a large fraction of proteins. Crucially, AlphaFold reports a per-residue confidence score, so users can tell which parts of a predicted structure to trust.

Why structure matters

A protein's shape determines what it does and how a drug might bind it. For decades, determining structures required slow, expensive experiments such as X-ray crystallography and cryo-EM, so most proteins had no known structure. AlphaFold's outputs are computational models — not experimental structures — but accurate enough to serve as useful starting points.

Analysis — the pattern we're watching

The single paper matters less than what it unleashed (analysis, not settled fact): within months, a public database released predicted structures for essentially the entire human proteome and hundreds of millions of proteins across species. The pattern is a shift from "structure is the bottleneck" to "structure is a commodity input," which moves the hard work elsewhere — onto validating predictions and onto the problems AlphaFold handles poorly. An active, still-unproven direction is using predicted structures to interpret disease-causing mutations and to design new binders; early results are encouraging, but each prediction remains a hypothesis to test in the lab.

What's still uncertain

Predictions are models, not proof. Confidence varies; disordered regions, large complexes, and protein motion are harder; and a predicted shape does not reveal function or prove a drug will bind. Experimental validation is still required before any clinical claim.