The problem. Predicting a protein’s 3D structure from its sequence was a 50-year open problem. Physics-based folding didn’t scale; earlier ML got close-ish. Then at CASP14, AlphaFold2 predicted structures at roughly experimental accuracy for most targets, and the field’s central question changed overnight.

The idea. Two coupled representations: a multiple-sequence-alignment track (evolutionary covariation, which residues mutate together, hinting at contacts) and a pairwise track over residue pairs, refined together in the Evoformer by repeated attention. A final structure module places atoms directly in 3D, end-to-end, and the whole thing recycles its own output to iterate. It emits a per-residue confidence (pLDDT) that turns out to be trustworthy, you can tell where to believe it.

Why it matters. For someone whose day job is analysis, the underappreciated angle is that AlphaFold2 is as much a data and evaluation win as a modeling one: decades of curated PDB structures, a ruthless blind benchmark (CASP), and a well-posed loss. The architecture matters, but the scaffolding around it is what made the result real and reproducible, the same lesson that governs any serious pipeline. Downstream, predicted structures and the confidence score are now infrastructure: inputs to design (ProteinMPNN, RFdiffusion), annotation, and drug work.

Verdict. Genuinely field-defining, and the caveats are well-known and honestly relevant: it predicts a static structure, leans on deep MSAs (weak for orphan/designed sequences), and says little about dynamics, complexes, or the effect of point mutations. Read it for the architecture, but steal the meta-lesson about data and evaluation. Pairs naturally with ESMFold as “with MSA vs. without.”