Implications of AlphaFold2 for crystallographic phasing by molecular replacement

Abstract

The AlphaFold2 results in the 14th edition of Critical Assessment of Structure Prediction (CASP14) showed that accurate (low root-mean-square deviation) in silico models of protein structure domains are on the horizon, whether or not the protein is related to known structures through high-coverage sequence similarity. As highly accurate models become available, generated by harnessing the power of correlated mutations and deep learning, one of the aspects of structural biology to be impacted will be methods of phasing in crystallography. Here, the data from CASP14 are used to explore the prospects for changes in phasing methods, and in particular to explore the prospects for molecular-replacement phasing using in silico models.

The X-ray crystallography phase problem solved thanks to AlphaFold and RoseTTAFold models: a case-study report

“…Like many other structural biologists (Cramer, 2021;Perrakis & Sixma, 2021;Thornton et al, 2021;McCoy et al, 2022;Subramaniam & Kleywegt, 2022), we are convinced that this achievement in protein structure-prediction accuracy will revolutionize structural biology. In the future, methods such as MIR/SIR or SAD/MAD will become increasingly marginal or applied to specific cases such as low-resolution data, for instance.…”

“…value lower than 1.5 A ˚with an already known protein crystal structure (Giorgetti et al, 2005;Scapin, 2013). The quality of the AlphaFold and RoseTTAFold models indicates that the 30% sequence identity threshold is no longer valid and that it is most important to have accurate models (McCoy et al, 2022). Several examples indicate that AlphaFold models exhibit r.m.s.d.…”

“…Regular progress in this field has been observed thanks to the Critical Assessment of protein Structure Prediction (CASP) biennial challenge (Kryshtafovych, Schwede et al, 2021), which was launched in 1994. Traditionally, template-based models were the most accurate (McCoy et al, 2022), although a few successful protein models were obtained by in silico modeling (Qian et al, 2007;Sjodt et al, 2018). For CASP13 (CASP round XIII; Kryshtafovych et al, 2019), the first version of the AlphaFold program (version 1.0; Senior et al, 2019), based on deep-learning and model-free approaches, led to 3D models with significantly improved quality (AlQuraishi, 2019).…”

“…As a rule of thumb, it was considered for a long time that the structure of a protein that shares at least 30% sequence identity could serve as a template to generate a search model (Abergel, 2013;Scapin, 2013). However, it is now obvious that beyond the sequence identity, the structural similarity between the search model and the crystal structure is most important (McCoy et al, 2022). Thanks to their high accuracy and also to the significantly improved estimation of the error in the coordinates, both AlphaFold and RoseTTAFold 3D structure models have already influenced the process of 3D protein structure determination by the molecular-replacement technique (Flower & Hurley, 2021;Kryshtafovych, Schwede et al, 2021;Milla ´n et al, 2021;Moi et al, 2021).…”

MrParse: finding homologues in the PDB and the EBI AlphaFold database for molecular replacement and more

“…CASP14 saw the stunning performance of Alpha-Fold2 (AF2), which in many cases produced predictions that resembled the target as closely as a different crystal form typically would . The value of predictions from AF2 and the AF2-inspired RoseTTAFold (Baek et al, 2021) as search models was quickly demonstrated, although some cases still required domain splitting or other editing (Milla ´n et al, 2021;Baek et al, 2021;McCoy et al, 2022;.…”