December 2020: DeepMind’s AlphaFold2 achieved a median CASP14 score above 92, making protein structures predictable at near-experimental accuracy.
2022: Meta AI’s ESMFold predicted protein structures from single sequences, at roughly 60 times AlphaFold2’s speed.
2022: ProteinMPNN used deep learning to design sequences for chosen protein backbones, reporting 52.4% sequence recovery versus Rosetta’s 32.9%.
2023: RFdiffusion generated new protein backbones and improved targeted-binder success by approximately tenfold over the previous state of the art.
2023: Deep-learning designs produced de novo luciferases whose best catalytic efficiencies approached those of natural luciferases, while organoid and custom-enzyme searches yielded source leads but no directly validated milestone in the retrieved evidence.
