Iterative structure prediction outperforms all-at-once structure prediction

Summary

Iterative protein structure prediction outperforms all-at-once structure prediction (1). This was pronounced for difficult targets (e.g., CASP14 and CASP15 proteins), whereas easy targets (CAMEO) could be predicted using either approach.

Figures

	Iterative / O(L³)			Argmax / O(L²)
Model	CAMEO	CASP14	CASP15	CAMEO	CASP14	CASP15
---	---	---	---	---	---	---
1.4B Open	0.830	0.705	0.733	0.805	0.640	0.677
1.4B Overtrained	0.846	0.714	0.750	0.825	0.651	0.700
1.4B	0.807	0.693	0.697	0.775	0.608	0.636
7B	0.870	0.742	0.764	0.852	0.607	0.726
98B	0.895	0.763	0.801	0.884	0.719	0.770
ESMFold	0.865	0.728	0.735
AlphaFold2	0.904	0.846	0.826

Ref (1)

1.

Hayes T, Rao R, Akin H, Sofroniew NJ, Oktay D, Lin Z, et al. Simulating 500 million years of evolution with a language model. Science. 2025;387(6736):850–8. Available from: https://doi.org/10.1126/science.ads0018