Predicting Deleterious Amino Acid Substitutions

Abstract

Identifying substitutions that affect protein function is of major interest for those studying proteins and their implications in disease. Disease-causing mutations tend to occur in structurally and functionally important sites, and a significant fraction of polymorphism sites are located in these regions (Sunyaev et al. 2000). It is estimated that each person is heterozygous for 24,000–40,000 amino acid-altering substitutions (Cargill et al. 1999). Predicting substitutions at these sites as deleterious or neutral may help identify disease-associated alleles. A recent single nucleotide polymorphism (SNP) study used an amino acid substitution scoring matrix, BLOSUM62, to classify each amino acid substitution caused by a SNP in a coding region as conservative or nonconservative (Cargill et al. 1999). However, use of a substitution scoring matrix may be inappropriate for predicting whether an amino acid substitution will affect a protein's function or structure because it generalizes and does not incorporate information specific to the protein of interest.

Substitution scoring matrices, such as BLOSUM62, have not been tested against experimental data for their ability to predict protein-altering substitutions. The BLOSUM62 matrix, like most matrices, is intended for database searching and pairwise alignment (Henikoff and Henikoff 1992), which is a different task than predicting deleterious substitutions. Substitution matrix scores are typically calculated from a log odds ratio of target frequencies, obtained by counting pairs of aligned amino acids, with the background frequencies of the amino acids. Substitutions to a more abundant amino acid have a lower score relative to a less abundant amino acid because the background frequency is lower for the less abundant amino acid. However, the overall abundance of an amino acid is irrelevant when considering whether an amino acid change is tolerated. On average, 14 out of the 19 possible substitutions for a given amino acid have negative scores from the BLOSUM62 matrix and are deemed nonconservative by Cargill et al. (1999). If nonconservative substitutions are predicted to be deleterious, then many substitutions will be predicted to affect phenotype. However, proteins actually contain many positions that have a high degree of plasticity in accommodating amino acid substitutions, as shown in previous mutagenesis studies (Bowie and Sauer 1989; Climie et al. 1990; Huang et al. 1992; Markiewicz et al. 1994). Therefore, experimentally testing all changes deemed nonconservative by a substitution matrix would be time-consuming and wasteful because of this overprediction, especially for large-scale studies such as examination of nonsynonymous SNPs (Lander 1996; Irizarry et al. 2000) or in genome-wide random mutagenesis projects (Bentley et al. 2000; Chen et al. 2000; McCallum et al. 2000).

Given a protein query, aligned sequences from the protein's family give position-specific information, which a substitution scoring matrix lacks. Residues that are conserved completely in the protein family are expected to be important for function, and even a conservative substitution at one of these residues may affect protein function. A substitution matrix may underestimate the severity of deleterious substitutions at these crucial positions. At some positions, any amino acid change can be tolerated in the protein if these positions are not involved in protein function or structure. Because these are expected to be neutral substitutions, one might expect amino acids in these positions of a protein alignment to be diverse. Therefore, the accuracy for predicting the phenotype that results from an amino acid substitution based on sequence alignment of protein family members should be better than using a generalized substitution scoring matrix.

SIFT is a sequence homology-based tool that sorts intolerant from tolerant amino acid substitutions and predicts whether an amino acid substitution at a particular position in a protein will have a phenotypic effect. SIFT predicts the phenotype resulting from a substitution more accurately than substitution scoring matrices for three data sets. In some exceptional cases, a substitution is predicted by SIFT to be neutral but experimentally does have a deleterious effect; these can be accounted for by query-specific interactions that are not conserved among the protein family members.

The effect of 4004 substitutions was assayed for LacI (Markiewicz et al. 1994; Pace et al. 1997), 336 substitutions for HIV-1 protease (Loeb et al. 1989), and 2015 substitutions for bacteriophage T4 lysozyme (Rennell et al. 1991). These three data sets are used to test prediction performance. Tolerant prediction accuracy is the number of substitutions correctly predicted to have no effect divided by the total number of substitutions that gave a wild-type phenotype under experimental test conditions. Subtracting the numerator from the denominator gives the number of substitutions that have been predicted to be deleterious but gave a wild-type phenotype under experimental conditions. Deleterious prediction accuracy is the number of substitutions correctly predicted to have an effect on the protein divided by the number of substitutions that affected protein. Subtracting the numerator from the denominator gives the number of substitutions that were predicted to have wild-type phenotype but gave a deleterious phenotype under experimental conditions. Total prediction accuracy is the total number of substitutions correctly predicted divided by the total number of substitutions. Experimental prediction accuracy is the number of substitutions that were experimentally shown to affect protein function divided by the number of substitutions predicted to affect function. For the biologist investigating substitutions predicted to have a deleterious effect, the experimental prediction accuracy reflects the proportion of predictions that will yield affected phenotypes experimentally.

Acknowledgments

Footnotes

REFERENCES