AlphaFold 2

By Satwik · February 18, 2026

AlphaFold 2 (Jumper et al., 2021) solved protein structure prediction to an accuracy that, for many proteins, rivals experimental methods. Given an amino-acid sequence it predicts the 3D structure, a problem that had resisted the field for decades. At the 2020 CASP14 assessment it outperformed all other approaches by a wide margin, and the 2021 paper and open-source release, followed by a public database of predicted structures, made the result usable by the whole scientific community.

Why it mattered

This is a landmark for machine learning applied to a hard scientific problem, distinct from the language-model story but part of the same era of scaled deep learning. Architecturally it was novel, an attention-based system, the Evoformer, that reasons jointly over evolutionary sequence alignments and spatial relationships, with an end-to-end structure module. It showed that carefully designed deep learning could deliver genuinely new scientific capability, not just incremental benchmark gains.

Reading angle

AlphaFold 2 is the clearest positive-impact case of the period, accelerating biology, drug discovery, and basic research. It also anchors serious dual-use discussion. Tools that predict and eventually help design biomolecular structures are broadly beneficial but sit near sensitive territory, and the biosecurity community has since debated where the lines should be for structure prediction and, more pointedly, for downstream protein design. Read AlphaFold as the example that AI's biggest wins and its thorniest governance questions can arrive in the same system, and as a case for building safety and access norms around scientific AI before, not after, capabilities mature.