Can AlphaFold2 predict the impact of missense mutations on structure?

By G. Buel, K. Walters. 2022

301 citations recorded by the source corpus.

To the Editor — Understanding the impact that missense mutations have on protein structure helps to reveal their biological effects. Although the structural prediction algorithm of AlphaFold2 is able to predict wild-type (WT) structures to high accuracy, it seems to fall short in predicting the impact of missense mutations on the three-dimensional (3D) structures of proteins. In a landmark achievement, the half-century goal of predicting protein structure from an amino-acid sequence was accomplished by AlphaFold2 (ref. 1), a deep-learning algorithm that relies on homology and experimental structures in the Protein Data Bank (PDB). The program and its predictions for 98.5% of the human proteome are publicly available2. However, one potential limitation is the insensitivity of AlphaFold2 to structure-disrupting mutations in an input sequence. This is because a database of structure-disrupting mutations does not exist, and AlphaFold2 therefore primarily bases its predictions on WT or homologous sequences instead. This drawback is important because missense mutations frequently associate with human diseases and single amino-acid mutations can lead to protein aggregation, misfolding and dysfunction. Being able to predict the effect of mutations of interest on the 3D structure of proteins will help structural biologists and non-structural biologists alike make informed hypotheses about their mechanisms of pathogenicity. Here, we comment on the implications of this AlphaFold2 limitation using three illustrative examples of selected domains for which experimental data for both WT and structure-disrupting mutations are available. The examples include ubiquitin-associated domains (UBAs) of a human Rad23 protein (hHR23a), BRCT (breast cancer 1 (BRCA1) C-terminal) repeats of BRCA1, and the actin motor protein Myosin VI MyUb domain. We provide comparisons of the AlphaFold2-predicted structural models of known mutants of these domains to their WT counterparts. The two UBAs of the proteasome shuttle factor and DNA-repair protein hHR23a contain a central leucine (L198 in UBA1 and L355 in UBA2) that is buried in the core3 and required for structural integrity4. Alanine substitution causes the UBAs to become intrinsically disordered;4 however, AlphaFold2 predicted alanine-substituted UBA1 (Fig. 1a) or UBA2 (not shown) to be structurally equivalent to WT UBA with only minor differences in the fold. We used the average root mean square deviation E1708 L1786 L1705 W1782

Subfields: Molecular Biology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Molecular Biology

Citations in the corpus

AlphaFold2 and its applications in the fields of biology and medicine (2023, background): “…predict the phenotypic effect of missense mutations, 209,210 it was also observed that the performance of missense mutations prediction is not good in other studies, 211 and there were only weak or no correlations between the output metrics of AF2 and changes in protein stability or functionality.”
Scaffolding protein functional sites using deep learning (2022, background): “…typically <2 Å and <1 Å respectively, with high model confidence (pLDDT > 80; Table S2); such recapitulation suggests the designed sequences encode the designed structures (although it should be noted that AF has limited ability to predict protein stability (25) or mutational effects (26, 27)).”
From sequence to Boltzmann weighted ensemble of structures with AlphaFold2-RAVE (2022, background): “The second limitation is that AF2 fails in predicting changes in protein structure due to missense mutations 2 .”
Targeting p53 pathways: mechanisms, structures and advances in therapy (2023, background): “AlphaFold2 is currently unable to correctly predict the structural impact of missense mutations.(330) Furthermore, the p53 protein functions by binding to various ligands such as DNA, small molecules, metal ions or other proteins.”
A structural biology community assessment of AlphaFold2 applications (2021, background): “However, some reports have indicated that AF2 does not appear to be well suited to predict the structures of mutated proteins 54,55 .”
AlphaFold2 can predict structural and phenotypic effects of single mutations (2022, background): “…another analysis, the authors appear to assume that structure-disrupting mutations should result in a large change in predicted structure or pLDDT. 13 This is plausible, but it is also possible that structure-disrupting mutations might alter the kinetic or thermal stability, with only minor…”
The accuracy of protein structures in solution determined by AlphaFold and NMR (2022, background): “Further support for this suggestion as the source of the difference comes from a recent paper (Buel and Walters, 2022), which shows that point mutations that introduce local unfolding may not be well predicted by AF2.”
Accurate proteome-wide missense variant effect prediction with AlphaMissense (2023, background): “Although AF is largely insensitive to input sequence variation and cannot accurately predict structural changes upon point mutation ( 26 ), we hypothesized that AF ’ s intrinsic understanding of multiple sequence alignments (MSAs) and protein structure provides a valuable starting point for models…”