AlphaFold2 Can Predict Single-Mutation Effects

AlphaFold2 (AF) is a promising tool, but is it accurate enough to predict single mutation effects? Here, we report that the localized structural deformation between protein pairs differing by only 1–3 mutations—as measured by the effective strain—is correlated across 3901 experimental and AF-predicted structures. Furthermore, analysis of $\sim 11000$ proteins shows that the local structural change correlates with various phenotypic changes. These findings suggest that AF can predict the range and magnitude of single-mutation effects on average, and we propose a method to improve precision of AF predictions and to indicate when predictions are unreliable.

Figure 1

(a) Local deformation per residue measured by effective strain, $S_i$, between wild-type (WT) and mutant (A71G) H-NOX protein, for experimental (orange) and AF-predicted (blue) structures. Dotted line indicates the mutated residue. (b) $S_i$ vs distance from the nearest mutated site, ${\delta }_{\mathrm{m}}$. (c) Comparison of $S_i$ obtained from experimental and predicted structures. (d) Overlaid WT (gray, 6BDD_A) and mutant (color, 6BDE_A), experimental (orange) and predicted (blue) structures. (e) Wild type protein with residues colored by $S_i$; location of A71G mutation is shown. (f) Distribution of $S_i$ between matched pairs of structures with the same sequence ($M=0$), for PDB, AF, and averaged AF ($⟨\mathrm{AF}⟩$) structures. (g) Distribution of correlation between PDB strain fields and equivalent fields from PDB, AF and DMPfold, shown for different numbers of mutations, $M$. (h) Residual correlation that is due to mutations, shown for the full dataset and a nonredundant version (NR); whiskers show bootstrapped 95% confidence intervals. (i) Correlation between PDB and $⟨\mathrm{AF}⟩$ strain fields, ${\mathbf{S}}_i^p$, across all pairs $p$ and residues $i$ that are within a distance ${\delta }_{\mathrm{m}}$ from a mutated site, shown for the full dataset and a nonredundant version (NR).

Figure 2

(a) Correlation (Pearson’s $r$) between blue fluorescence (mTagBFP2) and AF-predicted effective strain (ES), $S_i$, between WT and 8191 variants for all sequence positions $i$; positions of mutated residues are shown by dotted lines; chromophore site (Y65) is indicated (red circle). (b) Structure of BFP, with each residue colored according to $\text{Corr}(S_i,\mathrm{Fluor}.)$ (A); Y65 atoms are shown as spheres. (c) Strain at residue Y65 vs fluorescence for mTagBFP2 variants. (d) Fluorescence-strain correlation per residue vs distance from residue $i$ to Y65; mutated positions are indicated (blue circle). (e)–(h) Correlations between: $S_{A218}$ and red fluorescence (mKate2); $S_{L59}$ and green fluorescence (GFP); catalytic activity and $S$ at the active site (PafA); folding ability (fraction of active enzymes) and average strain, $⟨S_i⟩$, of the 50 residues that correlate best with folding ability (PafA).

Figure 3

Fluctuations are greater in flexible regions. Deformation (ES) between experimental hen lysozyme structures (194L_A and 6RTA_A), $S_i$ is correlated (Pearson’s $r$) with $B$ factor, relative solvent accessibility (RSA), and pLDDT (left). Distributions (kernel density estimates) of correlations for all proteins (right).

Figure 4

Correlations are higher if mutations have large effects in rigid regions. (a) Pearson’s correlation between PDB-AF $\mathbf{S}$-correlation and: mean and mutated residue values of flexibility ($B$ factor, RSA, pLDDT, $⟨S_i⟩$); fraction of secondary structure ($\alpha$-helix or $\beta$-sheet); magnitude of mutation effect (PDB-PDB $\mathbf{S}$ correlation, ES at mutated site in PDB and AF, $S_{\mathrm{m}}^{\mathrm{PDB}}$ and $S_{\mathrm{m}}^{\mathrm{AF}}$, BLOSUM score, frequency of mutation in MSA). Results are shown for the non redundant sample; whiskers show bootstrapped standard deviations.