Introduction

Determining if missense mutations are deleterious is critical for the analysis of genes implicated in disease. Mutation databases like the Breast Cancer Information Core (BIC) database (Szabo et al. 2000) are unable to determine the mutational effects of about 88% of BRCA1 (MIM# 113705) missense or non-synonymous single nucleotide polymorphism (nsSNP) mutations. Recently, various in silico methods that attempt to predict the phenotype of missense mutations have emerged (Chasman and Adams 2001; Ramensky et al. 2002; Herrgard et al. 2003; Krishnan and Westhead 2003; Mooney and Altman 2003; Ng and Henikoff 2003; Cai et al. 2004; Lau and Chasman 2004; Verzilli et al. 2005; Xiao et al. 2005). These include approaches that predict the degree of structural changes the mutation will cause and evolutionary approaches that identify and measure the degree of conservation. However, there may be discrepancies between different prediction methods (Tchernitchko et al. 2004). Besides, mutagenesis study is also an important approach for structure-based drug design (see, e.g., Chou 1993, 1996; Chou et al. 2003; Du et al. 2004, 2005a, b, 2007; Wei et al. 2005, 2006a, b, 2007; Zhang et al. 2006, 2007; Gao et al. 2007; Li et al. 2007a, b; Wang et al. 2007a, b, c, d; Wu and Yan 2007; and particularly some comprehensive reviews in this area: Chou 2004, 2006; Chou et al. 2006).

Both evolutionary and structural approaches have been used to classify the phenotype of missense mutations. Evolutionary approaches have different schemes measuring the degree of conservation in the multiple sequence-aligned regions. The majority of current approaches have not taken into account important aspects of an amino acid residue such as hydrophobicity, surface area, charge, and volume. Santibanez Koref and coworkers (Santibanez Koref et al. 2003) had proposed using amino acid properties but eventually adopted an approach (I20 parameter set) that did not take amino acid properties into account when assessing the effect of missense mutations. The I20 parameter set regards a conservative mutation to be as equally damaging or benign as a non-conservative mutation. In this paper, however, we investigated whether the use of amino acid properties, together with evolutionary distances and positional residue frequency, could improve the assessment of the degree of conservation in a set of evolutionary aligned amino acid residues.

We present here an evolutionary approach that is an extension of the method by Santibanez Koref and coworkers (2003) for the functional prediction of missense mutation. Santibanez Koref and coworkers (2003) used a phylogenetic approach to quantify the variability of an amino acid change and applied it to p53 (TP53; MIM#191170). Their results showed that a parameter that represents the identity of amino acids could be used to classify mutants with an accuracy of 95.6%. In our approach, benign substitutions were first classified based on the assumption that missense mutations that are maintained in orthologs are unlikely to affect the function of the protein. Then, taking into account the evolutionary distance between species and their relative selection pressure, a conservation score based on 16 amino acid properties was used to characterize the remaining substitutions. We tested our predictions on the experimentally verified set of bacteriophage T4 lysozyme annotations (Rennell et al. 1991) and applied the method to some of the most common missense mutations of the BRCA1 gene found in the BIC database. Our approach compares well with existing methods and may potentially be used to identify biochemical and structurally important amino acids and also be used to assess the significance of missense mutations documented in mutational databases.

Materials and methods

Analysis

Bacteriophage T4 lysozyme and human BRCA1 amino acid and nucleotide sequences were extracted from Genbank. Amino acid sequences were aligned using CLUSTALX 1.83 (Jeanmougin et al. 1998). The nucleotide sequences were aligned to amino acid sequences and were checked visually. Residues and their corresponding codons that created gaps in the reference sequences (T4 lysozyme or human BRCA1) were removed. Missense mutations that were maintained in any one of the aligned orthologous sequences were considered as benign (B). This criterion is also called “Deleterious Change Not Present” (DCNP) in the work by Santibanez Koref and coworkers (2003).

For mutations that were not found in any of the orthologous positions, a Z-score based on the work by Santibanez Koref and coworkers (2003) was calculated. Firstly, an S-score that calculates the degree of variability of a mutation is defined. This S-score takes into account amino acid properties such as hydrophobicity, molecular weight, volume of residues, etc., and compares their variation between the mutant residue m with all the other residues at a particular position in a set of aligned sequences. A high S-score means that a property (or a linear combination of properties) exists for which there is little variation among the residues in the wild-type sequences but a large difference between these and the mutant residue. The S-score can be defined as the maximum of the following expression adapted from Santibanez Koref and coworkers (2003)

$$ Q(e) = \frac{{{\left( {v^{*}_{m} - \overline{{v^{*} }} } \right)}^{2} }} {{1 + \frac{1} {{n + 1}}{\sum\nolimits_{x = 1}^n {{\left( {v^{*}_{x} - \overline{{v^{*} }} } \right)}^{2} } }}}. $$
(1)

Here, m is the mutant residue, x is a residue in the n aligned sequences, \( \ifmmode\expandafter\bar\else\expandafter\=\fi{v}^{*} = \tfrac{1} {n}{\sum\nolimits_{x = 1}^n {v^{*}_{x} } } \) and vector v * is the linear combination of amino acid properties that maximizes the expression. With a given nucleotide mutation rate and evolutionary distance from the reference sequence, a simulated alignment could be generated to represent amino acid residues that do not undergo evolutionary selective pressure. For each mutation, 1,000 of such simulations were performed and 1,000 S*-scores were generated from the simulated alignments. Since the missense mutants are defined with respect to the human wild type sequence, the observed human codon was used as the starting point to simulate the sequences from the remaining species (Santibanez Koref et al. 2003). Finally, for each mutation, a Z-score that computes the probability of the S-score being less than or equal to S*-score is defined. A small Z-score indicates a mutation with a higher tendency to affect the function of protein, and thus having the tendency to be deleterious. The detailed description of this numerical scheme can be found in the study by Santibanez Koref and coworkers (2003).

In our paper, the following modifications to the method described by Santibanez and coworkers (2003) were made. Z-scores were calculated from 16 sets of amino acid properties. Sets of properties used were hydropathy (Kyte and Doolittle 1982), hydrophilicity (Hpl) (Kuhn et al. 1995), mean polarity (Pol), accessible surface area (Sa) from (Radzicka et al. 1988), net charge (Chg) (Klein et al. 1984), coil formation (Coil) (Charton and Charton 1983), volume (Vol) (Grantham 1974), number of hydrogen bond donors (Hbd) (Fauchere et al. 1988), molecular weight (Mw) (Fasman 1976), isoelectric point (Ispt) (Zimmerman et al. 1968), normalized frequency of alpha-helix (Phelx), beta-sheet (Pbeta), beta-turn (Pturn) from (Chou and Fasman 1978) and PCI-III defined by (Sneath 1966) by principal component analysis of a large set of physicochemical properties. These properties were also used by (Santibanez Koref et al. 2003). All parameters were normalized prior to the calculations so that the variance across the 20 amino acids was 1. The 16 sets of Z-scores were summed up to classify the missense mutation by consensus (consensus score). Thresholds that gave the most accurate classification were used. Mutations with consensus score between 0 and 1.28 (with a limit of 0.08 for each of the 16 properties) were classified as deleterious (D). Mutations with a consensus score above 1.28 are classified as benign (B). From here forward, this summation of the 16 sets of Z-scores will be referred as the Property Consensus score.

(Rennell et al. 1991) designated four phases of phenotype for the size of bacteriophage plaques. “++” being similar to wild type, “+” being significantly smaller in size than wild type, “±” being small till the stage where discerning individual plaques become difficult, and “−” being no plaques produced at all. In our study, we simply considered “-” as deleterious, and “++” as benign. False positive rate (FP) was defined as the proportion of benign mutations that were incorrectly classified as deleterious. False negative rate (FN) was defined as the proportion of deleterious mutations that were incorrectly classified as benign. The overall accuracy was the percentage of the total predictions that were correctly predicted.

Sources of data

Bacteriophage T4 lysozyme

Lysozyme amino acid and nucleotide sequences from the following 33 strains were used: Bacillus amyloliquefaciens phage Morita2001 (AY030242.1), Bacillus phage phi 29 (X04962.1), Bacteriophage 933W (AF125520.1), Bacteriophage Aeh1 (AY266303.2), Bacteriophage APSE-1 (AF157835.1), Bacteriophage B103 (NC_004165.1), Bacteriophage K139 (NC_003313.1), Bacteriophage P1 (X87673.1), Bacteriophage phiKMV (NC_005045.1), Bacteriophage PZA (from B. subtilis) (M11813.1), Bartonella henselae str. Houston-1 (NC_005956.1), Bartonella quintana str. Toulouse (NC_005955.1), Coxiella burnetii RSA 493 (NC_002971.2), Dictyostelium discoideum (AC116986.2), Enterobacteria phage P22 (NC_002371.2), Enterobacteria phage RB69 (AY303349.1), Enterobacteria phage T4 (NC_000866.4), Escherichia coli K-12 MG1655 (U00096.2), Escherichia coli O157:H7 EDL933 (NC_002655.2), Helicobacter pylori strain TS142 (AY054410.1), Lactococcus bacteriophage phi-vML3 (X16178.1), Magnetospirillum magnetotacticum MS-1 Magn01_2730 (NZ_AAAP01002730.1), Methylococcus capsulatus str. Bath (AE017282.1), Nitrosomonas europaea ATCC 19718 (NC_004757.1), Pseudomonas aeruginosa phage PaP3 (AY078382.2), Psychrobacter sp. (NZ_AADI01000002.1), Salmonella enterica subsp., Enterica serovar Paratypi A str. ATCC9150 (NC_006511.1), Salmonella enterica subsp. Enterica serovar Typhi Ty2 (NC_004631.1), Salmonella typhimurium bacteriophage ES18 (X67137.1), Salmonella typhimurium bacteriophage ES18 (X67137.1), Salmonella typhimurium LT2 (NC_003197.1, AAL21602.1), Xylella fastidiosa Ann-1 Xfas001_28 (NZ_AAAM01000028.1) and Yersinia pestis CO92 (NC_003143.1). The regions spanning the codons 36–39, 61–74, 83–85, 162–164 had few representatives in the alignment and were excluded. Analysis was done from codon 2 to codon 161. To simulate the control sequences, a substitution rate of 4.5 × 10−9 per base per year (Ochman et al. 1999) was used with 1,000 simulations performed at each position (Santibanez Koref et al. 2003). Evolutionary distances were estimated with the program r8s (Sanderson 2003). The phylogenetic tree for the input of the r8s program was obtained from the Bootstrap Neighborhood-Joining method using CLUSTALX 1.83 package (Jeanmougin et al. 1998). The prediction results were compared with those of (Chasman and Adams 2001; Ng and Henikoff 2001; Cai et al. 2004) and also with the I20 parameter set of (Santibanez Koref et al. 2003).

Human BRCA1

Only codon region 225–1,365 of the human BRCA1 sequence is available for comparison with 55 other organisms. These organisms are from all 17 eutherian orders and one marsupial (metatherian). BRCA1 amino acid and nucleotide sequences from the following species were used: Aardvark (AF284030.1), American beaver (AF540622.1), Asiatic elephant (AF284022.1), black rhinoceros (AF284011.1), brown hare (AF284005.1), Norway rat (NM_012514.1), thick-tailed bush baby (AF019080.1), cat (AF284018.1), chimpanzee (AF019075.1), cow (NM_178573.1), Daubenton’s bat (AF203746.1), dog (U50709.1), woodland dormouse (AF332046.1), dugong (AF284019.1), elephant shrew (AF284029.1), European mole (AY121756.1), European shrew (AY057828.1), southern flying squirrel (AF284003.1), Malayan flying lemur (AF019081.1), fox squirrel (AF332044.1), giant anteater (AF484232.1), Hottentot golden mole (AF284027.1), gorilla (AF019076.1), gray dolphin (AY057825.1), greater glider (AY243455.1), northern gundi (AF540624.1), small Madagascar hedgehog (AF284025.1), western European hedgehog (AF284008.1), hippopotamus (AF284015.1), horse (AF284010.1), howler monkey (AF019079.1), human (U14680.1), hyrax (AF284024.1), Shaw’s jird (AF332048.1), red kangaroo (AF284033.1), koala (AY243445.1), llama (AF284012.1), Caribbean manatee (AF284020.1), mountain tapir (AY057830.1), house mouse (NM_009764.2), nine-banded armadillo (AF484222.1), orangutan (AF019077.1), pale-throated sloth (AF284002.1), pangolin (AF284009.1), pig (AF284014.1), Cape porcupine (AF540631.1), rhesus monkey (AF019078.1), Herbert river ringtail (AY243448.1), Cape rock hyrax (AF284023.1), shrew mole (AY121758.1), sperm whale (AF284016.1), springhare (AF540637.1), tailess tenrec (AF284026.1), large tree shrew (AF284006.1), common tube-nosed fruit bat (AF447502.1), common wombat (AF284031.1), and as outgroup, African clawed frog (AF416868.1) and chicken (NM 204169.1). The analyses were based on sequences between human codons 225 and 1,365. Sequence data that fully spans this region are available for 12 eutherian mammals. Codon numbers for all sequences are based on human sequence. Sequences start at codon 241 for fox squirrel, woodland dormouse, northern gundi, 260 for Shaw’s jird, 262 for springhare, 267 for Daubenton’s bat, 277 for common tube-nosed fruit bat, hyrax, 279 for shrew mole, European shrew, 280 for common wombat, 281 for European mole, 282 for greater glider, koala, Herbert river ringtail, 285 for western European hedgehog, 301 for brown hare, 316 for Cape porcupine and at 274 for the remaining 26 mammals. Sequences end at 609 for red kangaroo, 1,004 for greater glider, 1,032 for Herbert river ringtail, 1,144 for koala, 1,153 for common wombat, 1,197 for common tube-nosed fruit bat, 1,201 for brown hare, 1,211 for northern gundi, 1,212 for small Madagascar hedgehog, 1,213 for shrew mole, European shrew, European mole, elephant shrew, 1,214 for Daubenton’s bat, 1,215 for Asiatic elephant, fox squirrel, 1,218 for hyrax, 1,222 for Cape porcupine, 1,338 for Shaw’s jird, springhare, 1,344 for American beaver and at 1,219 for the remaining 22 mammals. To simulate the control sequences, a substitution rate of 2 × 10−9 per base per year was used, with 1,000 simulations performed at each position. Evolutionary distances were estimated as described by (Kumar and Hedges 1998). The DCNP (Deleterious Change Not Present) criterion which assumes that a mutation leading to a residue being present at an orthologous position as being benign, was compared. For mutations that were not classified as benign by the DCNP criterion, the Property Consensus score was used to assess the effect of the mutation. This procedure that incorporates both the DCNP criterion and the Property Consensus score is called the Property Consensus with DCNP from here onwards (Fig. 1). Predictions derived from the Sorting Intolerant From Tolerant (SIFT, version 2) (Ng and Henikoff 2002) and Polymorphism Phenotyping (PolyPhen) (Ramensky et al. 2002) programs were also compared.

Fig. 1
figure 1

Flowchart showing the steps involved in the prediction of the property consensus with DCNP method

Full size image

Analysis of predictions of T4 lysozyme

T4 lysozyme structures 2LZM (Weaver and Matthews 1987) and 1L10 (double-mutant T155A, T157I) were downloaded from Protein Data Bank. Structures were viewed with Swiss-PDBViewer (Guex and Peitsch 1997). T4 lysozyme residues that our approach predicted to be deleterious, four or more times, for the 13 substitutions conducted by (Rennell et al. 1991) are listed in Table 2 (see below). The Perf Software (http://kodiak.cs.cornell.edu/kddcup/software.html) was used to derive the Receiver Operator Characteristic Plot (ROC) for our prediction and the I20 parameter prediction.

Results

Evaluation of the property consensus with DCNP method by comparison with other prediction methods

Figure 1 describes the prediction process for a given missense mutation. The first step in the prediction process is to check if the mutation has been maintained in the corresponding position in the aligned orthologous sequences. Mutation that is not maintained in the orthologs will be subjected to the calculation of Z-scores for the prediction of its effect on the protein.

The approach used in this study is called the Property Consensus with DCNP method because it predicts the effect of a mutation based on the consensus of several amino acid properties. When predicting the effect of a mutation, each of the properties examined has an equal weight in deciding the outcome of the prediction. In order to evaluate the accuracy of the Property Consensus with DCNP method, experimentally verified T4 lysozyme mutations were analyzed with this method and other computationally derived prediction methods (Table 1). Predictions were compared in terms of the percentages of FP, FN and overall accuracy.

Table 1 Comparison of prediction methods for evaluating missense mutations using the T4 lysozyme data sets
Full size table

Table 1 shows that the Property Consensus with DCNP method, in terms of the overall accuracy, ranked first out of the five methods. The high false positive rate (60%) of the Property Consensus with DCNP approach was possibly due to an over-estimation of benign mutations, resulting from a high representation of biochemical properties among the 16 properties favoring benign prediction. When a threshold of 0.01 is used for the I20 parameter method by (Santibanez Koref et al. 2003), a relatively high false positive rate of 84.2% was obtained. However, it has also one of the lowest false negative rate of 5.8%. SIFT prediction also had a relatively high false positive rate of 41% (Ng and Henikoff 2001). Though Cai et al. (2004) had the lowest false positive rate of 2% among the approaches compared, their method had the highest false negative rate of 72%. From Fig. 2, the Area Under the ROC curve (AUC) for Property Consensus with DCNP has a similar AUC to that for the Property Consensus without DCNP (AUC of 0.7). In contrast, the AUC for the I20 parameter (0.36) is considerably lower than that for the Property Consensus methods

Fig. 2
figure 2

ROC curves for the T4 lysozyme dataset. The area under the curve (AUC) for the Property Consensus with DCNP method (blue line, represented by DCNP in the legend) is 0.702. The AUC for Property Consensus without DCNP method (pink line, represented by “without DCNP” in the legend) is 0.706. The AUC for the I20 parameter set (green line, represented by “I20” in the legend) is 0.359

Full size image

Analysis of the property consensus with DCNP method on T4 lysozyme mutations

In the study by (Rennell et al. 1991), each amino acid in the T4 lysozyme was substituted to 13 other amino acids. Table 2 lists all the T4 lysozyme residues where the Property Consensus with DCNP method predicted them to be deleterious, four or more times, for the 13 substitutions (Rennell et al. 1991).

Table 2 List of all the T4 lysozyme residues that the property consensus with DCNP method predicted to be deleterious, four or more times, for the 13 substitutions conducted by (Rennell et al. 1991)
Full size table

These residues therefore represent amino acids that may play important structural and biochemical roles in the protein, concurring with data from published studies (Alber et al. 1987; Grutter et al. 1987; Matthews 1996; Goto et al. 2001). These include mutations at putative active sites (Glu 11, Tyr 18, Thr 26), sites creating loop turns (Gly 28, Gly 30, Gly 107), buried hydrophobic residues (Leu 7, Leu 99), polar exposed residues that may be involved in dipolar coupling (Thr 142) and which may affect the thermal-stability of the protein (Thr 151, Thr 155). A structural representation of these mutations is illustrated in Fig. 3.

Fig. 3
figure 3

Structure of T4 lysozyme (2LZM) in coil representation colored by secondary structure (red helices, yellow sheets). The SwissPDB-Viewer (Guex and Peitsch 1997) has been used to prepare these cartoons. Residues that the Property Consensus with DCNP approach predicted to be deleterious in more than 3 out of 13 mutations conducted by (Rennell et al. 1991) are represented. Conserved glycine residues are colored in green (28, 30, 107) forming loop turns, residues found in the active site cleft are colored in magenta (Glu 11, Tyr 18, Thr 26), buried leucine residues are in yellow (7, 99) with Leu 99 pointing into a cavity (represented as the blue globular structure), threonine residues at the entrance to the active site are in light red and threonine residues near the C-terminal are colored in orange

Full size image

All the eleven sites from Table 2 are represented in Fig. 3. The probable importance of Thr 151 and Thr 155 in maintaining structural stability is illustrated in Fig. 4 with reference to (Alber et al. 1987; Grutter et al. 1987).

Fig. 4
figure 4

Structure of residues 151–159 of 2LZM illustrating the effect of a T155A on hydrogen-bonding. a Residues 151–159 of wild type T4 lysozyme (PDB code: 2LZM) (Weaver and Matthews 1987). The gamma-hydroxyl group of Thr 155 (red) forms a hydrogen bond with Thr 157 (yellow). Asp 159 is colored in blue and Thr 151 in magenta. b Residues 151–159 of the double-mutant (T155A, T157I) T4 lysozyme (PDB code: 1L10). A mutation of Thr 155 to Ala 155 results in loss of hydrogen-bonding between residues 151 and 157

Full size image

Analysis of various prediction methods on BRCA1 mutations

Twelve of the most frequent missense mutations listed in BIC between codons 225 and 1,365 were subjected to the various prediction methods (Table 3). The remaining eight most frequent missense mutations do not fall in this region of Brca1. The PolyPhen program predicted two possibly damaging (pD) and two probably damaging (D) mutations. The property consensus scores that are below 5.0 could be classified as intermediate (I) or possibly deleterious predictions. With such a classification, the prediction matches exactly with those predicted by PolyPhen (R1347G, R841W, Q356R, N550H). The prediction given by SIFT however, matches only marginally with PolyPhen or the Property Consensus with DCNP method (Tchernitchko et al. 2004).

Table 3 Analysis of the 12 most common missense mutations reported in breast cancer information core (BIC), in the region of codon 225 to 1,365, in descending order of frequency as recorded in BIC
Full size table

Discussion

Comparison of the property consensus with DCNP method to other methods

Evaluation of the Property Consensus with DCNP method using the T4 lysozyme mutation data set (n = 1,366) had an overall accuracy of 87.6% when a threshold of 1.28 (0.08 for each of the 16 properties) was used (Table 1). There are limitations when comparing prediction methods based on percentages of false positives rates, false negatives rates and overall accuracy generated from different methods and mutational datasets. For example, if a dataset such as T4 lysozyme contains mutations that are largely tolerant with high degree of plasticity (Matthews 1996), a prediction method that has a tendency to over-predict tolerant mutations may still have a high overall accuracy. On the other hand, if the dataset went through a series of selection criteria to include only the important and well conserved residues, and when the prediction method tends to over-predict deleterious mutations, such a prediction will also have a relatively high overall accuracy.

Intuition favors the assumption that a mutation is most likely to be benign if it is found to be maintained in the same position in an alignment of orthologs. However, the Property Consensus with DCNP method did not produce a considerably better prediction than the Property Consensus method without DCNP (Fig. 2; both with AUC of 0.7). This phenomenon can be partially explained by the relatively high rate of co-evolution of residues within the bacteriophage sequence. For example, codon 10 of T4 lysozyme is an Aspartate residue and many other species has a Tyrosine maintained at codon 10. Hence, the Property Consensus with DCNP method will treat D10Y as a benign mutation. However, due to co-evolution of interacting residues, where the Aspartate at codon 10 has an association with the Tyrosine residue at codon 161 (an Aspartate-Tyrosine association), the other species have Glycine maintained at 161 (i.e., a Tyrosine–Glycine association). So, in this case, a single mutation from Aspartate to Tyrosine (Tyrosine–Tyrosine association) will result in steric hinderance of two bulky side-chain residues that will cause disturbance to the structure of T4 lysozyme. Such instances will increase the rate of false negative of the prediction.

The I20 parameter prediction initially proposed by (Santibanez Koref et al. 2003), which did not consider amino acid properties, produced an overall accuracy of 44.6% and AUC of 0.359. In the I20 parameter prediction, a conservative substitution from an Aspartate to Glutamate is considered equally benign or equally deleterious when compared to a non-conservative substitution from an Aspartate to Tryptophan, independent of the position of the residue. Hence, the relatively high false positive rates could be due to its tendency of predicting a benign substitution as a deleterious substitution. In the study by (Santibanez Koref et al. 2003) on TP53, the assessment of the performance of the I20 parameter was based on a set of controls (defined as amino acid residues that could be found in the corresponding position of the aligned ortholog sequences) and mutants [defined as missense mutation found in the TP53 database (Beroud and Soussi 1998) but were not found in the corresponding position of the aligned ortholog sequences]. From our analysis, we found that this assumption, which their study took as the set of controls, does not always hold true.

Moreover, other studies (Rennell et al. 1991; Matthews 1996) have shown that not all missense mutations, which do not fit their definition of controls, would truly affect the function of a protein. Hence, the relatively good performance of the I20 parameter in their study on TP53 may, in part, be due to the possibility that TP53 may be more sensitive to mutational changes as compared to proteins like T4 lysozyme that exhibits a higher tolerance to mutational effects (Rennell et al. 1991; Matthews 1996). However, the relatively good performance of the I20 parameter may also be due to their definition of deleterious mutants which, again, would have the tendency of treating benign mutations as deleterious.

When the Property Consensus with DCNP method was compared to the SIFT method (Ng and Henikoff 2001), it had a higher overall accuracy and comparable false positive and false negative rates. Although our prediction method could be applied to all proteins with sufficient number of orthologous sequences, its accuracy has not been proven in distant proteins of a different function or structural stability. However, these evolutionary-based prediction methods may be expected to predict mutations with a similar level of accuracy when they are applied to subfamilies of proteins with similarity in their structure and function.

Structural and biochemical evidence that support the property consensus with DCNP method’s predictions of functionally important residues in T4 lysozyme

Table 2 lists all the residues in T4 lysozyme that the Property Consensus with DCNP method predicted to be deleterious, four or more times, for the 13 substitutions conducted by (Rennell et al. 1991). Most of these sites have been known for their structural and functional importance. The T4 lysozyme family has an important triad in the active site comprising of Glu 11, Asp 20 and Thr 26 (Kuroki et al. 1993). Included in the list in Table 2 was Glu 11 and Thr 26 but not Asp 20, which may be the least important residue in the triad as studies have shown that catalytic activities could still be carried out when Asp 20 is replaced by some other amino acids (Rennell et al. 1991; Kuroki et al. 1999). All 12 substitutions of Glu 11 were deleterious and 11 out of the 13 substitutions of Thr 26 were deleterious, while in contrast, only 9 out of 13 substitution of Asp 20 were deleterious (Rennell et al. 1991). Tyr 18, which is also found in the active site cleft of T4 lysozyme, is, however, replaceable by several other amino acids. Tyr 18 was included in the list in Table 2 as one of the important residues because it is conserved in 31 out of the 33 organisms used in this study. Thus, the Property Consensus with DCNP method was able to accurately predict biochemically important residues in T4 lysozyme.

Gly 28, Gly 30 and Gly 107 are well-conserved residues responsible for the loop turns as illustrated in Fig. 3. Leu 7 and Leu 99 are buried hydrophobic amino acids and their substitutions by hydrophobic residues will most likely disrupt its structure. Leu 99 extends into a relatively large cavity within the lysozyme structure (Fig. 3). Substitutions on Leu 99 that produce an even larger cavity within the protein tend to be very destabilizing (Matthews 1996). The list in Table 2 also shows Thr 142 to be an important site.

Replacement of Thr 21 and Thr 142 to a disulphide linkage abolishes the activity of T4 lysozyme, because this link bridges across the active-site cleft, preventing the substrate from entering the active site of the enzyme (Matsumura et al. 1989). Hence, substitutions on Thr 21 or Thr 142 by charged or bulky residues may cause lysozyme to lose its catalytic activity. The experimental study by (Rennell et al. 1991) showed that substitution of Thr 142 by an Arg or Lys had a deleterious effect on the enzyme activity. Substitutions of Thr 151 and 155 disrupt part of a network of hydrogen bonds (surrounding residue Thr 157), which may result in unsatisfactory hydrogen-bonding potential that could possibly contribute to a reduction of stability (Table 2, Fig. 4). (Grutter et al. 1987) showed experimentally that disruptions to this network of hydrogen bonds were indirectly responsible for reducing the temperature of the midpoint of the reversible thermal denaturation transition by 11°C. The identification of these residues demonstrates that the Property Consensus with DCNP method was able to accurately predict structurally important residues in T4 lysozyme.

A sequence-based evolutionary approach such as the Property Consensus with the DCNP method can produce structural inferences because amino acid properties used in the study contain parameters that directly or indirectly affect structural stability. The ability of this method to sieve out structurally and experimentally important residues from a multitude of substitutions shows that this approach may be potentially useful in the discovery of biochemical and structurally important residues.

Comparison of predictions for twelve frequently occurring BRCA1 missense mutations

Previous structure-based studies on missense mutations in BRCA1 have looked at the BRCT domains (codon region 1,640–1,729, 1,760–1,821) at the C-terminal of the protein (Williams and Glover 2003; Mirkovic et al. 2004). However, there are only a few orthologous sequences available in this region and hence this region is not viable for a sequence-comparison method such as the approach used in this study. Hence, this study can only examine the region 225–1,365 where more orthologous sequences can be found. As such, our BRCA1 predictions can only be compared to prediction methods that can examine missense mutations within this region.

Missense mutations that occur frequently may play a significant role in disease penetrance in a given population. Of the 20 most frequently occurring missense mutations recorded in the BIC BRCA1 database, 12 of them fall in the BRCA1 225–1,365 region and were analyzed with various approaches (Table 3). SIFT predicted 5 deleterious mutations (D) out of the 12 missense mutations. Only 2 of these predicted deleterious mutations overlapped with the deleterious mutations predicted by PolyPhen and the Property Consensus with DCNP method. On the other hand, all of PolyPhen’s 4 possibly deleterious (pD) and D mutations were also predicted as being possibly deleterious by Property Consensus with DCNP.

Discrepancies between SIFT and PolyPhen have previously been documented, and it has been suggested that PolyPhen may perform better because, in addition to sequence information, it incorporates structural information of the protein in its prediction (Tchernitchko et al. 2004; Johnson et al. 2005). Notably, although PolyPhen’s methodology differs from that of the Property Consensus with DCNP method, their prediction results were similar. This similarity in the prediction between PolyPhen and Property Consensus with DCNP could partially be attributed to the latter’s incorporation of structurally related amino acid properties. Hence, results from our study are in line with previous inferences that PolyPhen may outperform SIFT in predicting the phenotype of nsSNPs. PolyPhen (polymorphism phenotyping) developed by (Ramensky et al. 2002) used empirically derived rules to assess a set of analysis that calculates the PSIC (position specific independent counts) (Sunyaev et al. 1999) profile scores for two amino acid variants. It then combines information on sequence features, the structural parameters and contacts to characterize the substitution. In line with results from previous studies (Barker et al. 1996; Durocher et al. 1996; Dong et al. 1998; Petersen et al. 1998; Fleming et al. 2003), our results also show that the BRCA1 missense mutations R1347G, R841W and Q356R with a population frequencies of 4.3, 3.2 and 2.3%, respectively, are candidate mutations that may play important roles in assessing cancer susceptibility in a population.

Factors that affect the performance of the property consensus with DCNP method

We highlight here some of the factors that may affect the accuracy of prediction by the Property Consensus with DCNP method. Firstly, the level of confidence of its prediction will largely rely on the availability of sequences in other species. Regions where only few hits could be obtained from homology searches (in this case, the region before codon 225 and after 1,365 of BRCA1) will decrease the level of confidence in the prediction. However, as more orthologous genes and genomes are being sequenced, this limitation will eventually be eliminated. Secondly, changing the assumed rate of mutation by one or two orders does not substantially affect the prediction (Santibanez Koref et al. 2003). In the calculation of the predictive score, the degree of conservation between orthologs is considered as a more important factor than the degree of conservation between the controls (i.e., the simulated sequences representing genes that do not evolve under selective pressure). Thirdly, depending on the function of the protein (e.g., structural protein or enzymatic protein) and the site of the mutation in the three-dimensional conformation of the protein, the importance of amino acid properties will vary. For example, we may expect the hydrophobic character of a residue to be more important for a buried site and less important at the protein surface. Giving equal weight to all the amino acid properties may not adequately represent their in vivo conditions. Fourthly, co-evolution of binding partners or acquisition of compensatory mutations (Kondrashov et al. 2002) is also likely to increase with divergence times and higher mutation rates. Hence, the DCNP criterion may not be a good measure to sieve out benign phenotypes especially for distant orthologs or sequences with high mutation rates. Fifthly, this study has compared predictions only with the experimentally verified T4 lysozyme substitutions. Thus, its prediction accuracy for a protein from different functional or structural subclasses is not known. However, it may still be used to retrieve preliminary information of a protein but should not be used in the absence of other tests or arguments to reach conclusions for clinical assessment.

Conclusions and future work

This study shows evidence that the Property Consensus approach has the potential to provide relatively accurate predictions in phenotyping missense mutations. The use of an orthogonal set of amino acid properties (Kidera et al. 1985a, b) instead of choosing some important properties or using all of the 500 available amino acid properties (Kawashima et al. 1999) may more closely model functionally important aspects of amino acid residues. In addition, different weights could be assigned to different properties by conducting a sequence-based characterization prior to the prediction process. For example, trans-membrane, signal region, or binding, active site predictions, based on sequence-based characterization at the region surrounding the site of substitution could be conducted, as in the case of PolyPhen. Statistical methods that measure the relative importance of various amino acid properties at a given site of substitution could also be explored. More experimentally verified data from diverse sets of proteins are necessary to understand the relative importance of this orthogonal set of amino acid properties in their respective characteristic sites.

In summary, we have shown that the use of amino acid properties (Kawashima et al. 1999) has good potential in improving the measurement of the degree of conservation, and hence the degree of importance of a point missense mutation. It may also be used by scientists who want to identify functionally important residues in a protein. Such functional prediction methods may contribute to the improvement of functional annotation of mutational and SNP databases and ultimately may improve our understanding of the role of disease susceptibility and various SNP-related complex diseases.