Stabilizing proteins from sequence statistics: the interplay of conservation and correlation in triosephosphate isomerase stability

Abstract

Understanding the determinants of protein stability remains one of protein science's greatest challenges. There are still no computational solutions that calculate the stability effects of even point mutations with sufficient reliability for practical use. Amino acid substitutions rarely increase the stability of native proteins; hence, large libraries and high-throughput screens or selections are needed to stabilize proteins using directed evolution. Consensus mutations have proven effective for increasing stability, but these mutations are successful only about half the time. We set out to understand why some consensus mutations fail to stabilize, and what criteria might be useful to predict stabilization more accurately. Overall, consensus mutations at more conserved positions were more likely to be stabilizing in our model, triosephosphate isomerase (TIM) from Saccharomyces cerevisiae. However, positions coupled to other sites were more likely not to stabilize upon mutation. Destabilizing mutations could be removed both by removing sites with high statistical correlations to other positions and by removing nearly invariant positions at which "hidden correlations" can occur. Application of these rules resulted in identification of stabilizing mutations in 9 out of 10 positions, and amalgamation of all predicted stabilizing positions resulted in the most stable yeast TIM variant we produced (+8 °C). In contrast, a multimutant with 14 mutations each found to stabilize TIM independently was destabilized by 2 °C. Our results are a practical extension to the consensus concept of protein stabilization, and they further suggest the importance of positional independence in the mechanism of consensus stabilization.

Fig. 1

re-S.c. TIM. (a) Histogram of relative entropy values for all 240 aligned positions in the TIM family. The mean RE is 1.42. (b) The six most conserved positions in S.c. TIM that are not consensus amino acids are shown in green sticks. The active-site residues are shown in orange on the 1YPI crystal structure. (c) Ellipticity at 222 nm is followed with increasing temperature. The wild type melts at 59.1 °C, but re-S.c. TIM melts at 57.0 °C.

Fig. 2

CD characterization of highly conserved mutations. (a) The circular dichroism spectra of wild type and consensus variants of TIM. All have similar ellipticity when normalized for concentration except W90Y, which may be partially unfolded. (b) The CD thermal melts indicate that four individual consensus variants are more stable than wild type, but the remaining two are less stable.

Fig. 3

Thermal stabilities of consensus TIM variants. We monitored the loss of secondary structure with increasing temperature at 222 nm for helices (a) and 215 nm for sheets (b). (c) The optical density at 600 nm from aggregation reports similar two-state unfolding profiles as the CD thermal melts. (d) High-throughput thermal scanning was used to assay the melting temperatures based on hydrophobic dye binding. Note that the same colors are used for the same variants in (a)–(d).

Fig. 4

Concordance of stability assays. The variants are arranged by the T_1/2 values derived from CD thermal denaturation at 222 nm. Data were not collected by high-throughput thermal scanning for A66C, I109V, D180Q, and A212V.

Fig. 5

Filtering by conservation. (a) All positions in TIM have been plotted against their relative entropies from the neutral reference state. All sites are shown in gray, and stabilizing and destabilizing consensus mutations are shown in green and red, respectively. Note that the stable mutations aggregate above the black arrow, which indicates the mean relative entropy of 1.42. (b) Amino acid distributions for the yeast neutral reference state and positions with relative entropy values of 0.5, 1.0, 1.5, 2.0, and 2.5 are shown.

Fig. 6

Mutual information and protein stability. (a) The mutual information matrix for all 240 positions in TIM is shown. The matrix is symmetric (x–y is the same as y–x), and there is no meaning to the self-correlations (x–x), which were not calculated. (b) The distribution of mutual information scores is shown for the entire matrix. Here, approximately 30% of all pairwise correlations are above the noise threshold of 0.23. The distribution of mutual information scores are shown for stabilizing mutations (c) and destabilizing mutations (d). Note that there is a significantly higher fraction of strong correlations at the positions that lead to a loss in stability.

Fig. 7

A hidden correlation between positions 11 and 20. (a) The crystal structures of S.c. TIM and T. maritima TIM [Protein Data Bank entries 1YPI (pink) and 1B9B (green)] are aligned and residues 11 and 20 are highlighted. The F11W mutation may have introduced a steric clash resulting in destabilization. (b) CD thermal denaturation of F11W and F11W I20A. I20A alone did not express in appreciable quantities.

Fig. 8

Characterization of algoTIM and comboTIM. (a) The CD wavelength scans of wild-type S.c. TIM and algoTIM are nearly identical. comboTIM shows less ellipticity at 222 nm and has its deepest minima at 205 nm, suggesting some random coil. (b) The CD thermal melts monitored at 222 nm are shown for all characterized proteins in gray, with comboTIM, S.c. TIM, the F11W I20A mutant, and algoTIM highlighted.

Fig. 9

Mutual information for comboTIM and algoTIM mutation sites. The positions of mutation for comboTIM and algoTIM have been isolated from the mutual information matrix of all pairwise interactions. (a) The positions of mutation in algoTIM have virtually no strong (red, orange) correlations to other sites in the protein. (b) In contrast, the 14 positions of mutation in comboTIM have many strong correlations with other positions within TIM. (c) The 15 mutations in algoTIM are assembled into a matrix with the correlations displayed as a heat map. The positions of mutation are not correlated to each other. (d) The 14 mutations in comboTIM are assembled into a matrix with the correlations displayed as a heat map. Although these mutations were stabilizing independently, there are many strong correlations between sites of mutation in comboTIM, perhaps leading to nonadditive effects.