Hi, this is Yitian. I'm documenting my thinking as I learn AIDD.
Biohub and EvolutionaryScale's new paper chains ESMC, ESMFold2, therapeutic binder design, sparse-autoencoder interpretability and a 6.8-billion-sequence atlas into one argument about emergence in protein language models. Reading the chain, and where it stops.
BindCraft proved that back-propagating through AlphaFold works. Four follow-ups — BindEnergyCraft, FoldCraft, Germinal and mBER — each repair a different gap, and together they argue that the field's real contest is about where to put the constraints.
AlphaFold3-class predictors hallucinate plausible structures even from all-X sequences. Protein Hunter turns that failure mode into a zero-shot design loop: hallucinate a structure, inverse-fold a sequence, re-predict, repeat.
RFdiffusion went from residue-level backbone generation to all-atom interaction design in about two years. Reading the RFD1, RFD2 and RFD3 papers as one arc, and what each generation actually changed.
Notes on the Structure review "Code to complex: AI-driven de novo binder design" — forty years from rational design to generative design, the four current design philosophies, and where the bench-to-bedside stories already are.
BindCraft and its relatives turned AlphaFold2 into a scoring function and back-propagated through it to hallucinate binders. A survey of how that idea works, what it buys us, and where it still breaks.
