Integration of curated databases to identify genotype-phenotype associations

Prediction of genes associated to phenotypes

After analyzing the accuracy of the data sets, it is also possible to make reasonable hypotheses for COG-phenotype pairs that are characterized but have not yet been confirmed by the biological literature. These COG-phenotype pairs are listed using their GIDEON identifier/description-COG description/protein name. One example is the B31/Catalase-COG1651/Protein-disulfide isomerase (DsbG) pair with a correlation score of 0.91. Dsb proteins are known to oxidize the sulfhydryl groups of periplasmic proteins to disulfide bonds, donating electrons to ubiquinone, and thereby making the electron transport chain the primary source of oxidizing power for sustaining periplasmic sulfhydryl oxidation [19, 20]. During the stationary phase, electron transport to oxygen is reduced. Bandyopadhyay et al. suggest a possible complementary role between catalase and the Dsb proteins in maintaining periplasmic sulfhydryl oxidation. It is possible that catalase may be critical in peroxidatically oxidizing ubiquinol or another periplasmic or inner membrane component using H₂O₂ as an electron acceptor during the stationary phase when the oxidizing capacity of the electron transport is diminished [21].

With a correlation score of 0.95, other possible associations can be made for the FAP/L-Rhamnose phenotype with various phosphotransferase system sorbitol-specific component proteins. Some microbes such as the Klebsiella I-174 organism make exopolysaccharides with a high rhamnose content [22]. Farres et al. showed that the addition of sorbitol increased the production and growth of rhamnose over other carbon sources such as sucrose [23]. This study suggests that proteins involved with sorbitol metabolism and utilization could be linked to rhamnose production.

Mapping organisms between databases

The laboratory results in the GIDEON database are primarily used for identifying bacterial species for medical diagnostics. Since different strains of the same bacterial species are often sequenced, NCBI's taxonomic annotation is sometimes at the subspecies level. In contrast, the GIDEON phenotypes do not achieve such a high resolution, and for this reason GIDEON taxonomic annotation is established at the level where the phenotype is consistent in all descendants of the phylogenetic tree (generally the species level).

This presented a complication in integrating the two data sources. To overcome this, we assumed that phenotypes from the microbiological database for one species are valid for every subsumed subspecies and strain listed in the COGs database. This is a valid assumption since the GIDEON dataset provides microbiologists with relevant tests designed to distinguish between organisms according to their phenotypes. Thus if the phenotype is specific to a subspecies, it will be annotated at the level of the subspecies, but if the phenotype is common to all subspecies, it is recorded at the level of the species.

Following this principle, we first identified the taxonomic level for the fully sequenced bacteria in the COGs dataset, and then used text string matching followed by manual examination to map the species in GIDEON [24]. As a result, we have mapped the 37 microorganisms present in both GIDEON and the COGs. Of the 37 mappings, 23 have identical species annotations in GIDEON and COGs, and 9 have a species annotation in GIDEON mapped to one or more subspecies in COGs.

There were several COGs species including H. pylori, E. coli, M. tuberculosis, and N. meningitidis which had more than one subspecies with complete genome sequences. In these cases, the subspecies were merged to the single GIDEON species by selecting only the COGs common to all subspecies. In this manner, we eliminated the subspecies specific differences that the phenotypic assays would have been unable to resolve. We generated a matrix showing the presence and absence of COGs across these 37 species.

Associating genes to phenotypes

We employed a correlation analysis to quantify the association between a given COG and a GIDEON phenotype. Two matrices were constructed. We defined X as a two-dimensional matrix indicating the presence or absence of organisms within a COG (X was constructed as an M × N matrix, where M is equal to the number of COGs and N is equal to the number of organisms within a COG). For the corresponding GIDEON lab conditions, a similar distance matrix, Y, was constructed as an N × L matrix, where N is equal to the percent survival of organisms subjected to each lab condition and L is equal to the number of lab conditions. X _ij is the presence or absence of a COG m _i within an organism n _j , and Y _jk signifies the percent response of an organism n _j under a certain lab condition l _k . We computed an M × L matrix of linear correlation coefficients r _ik (Pearson's correlation coefficient [25], where each r _ik is defined as:

$r_{ik}=\frac{{\displaystyle \sum _{j=1}^N\left(X_{ij}-{\overline{X}}_i\right)}\left(Y_{jk}-{\overline{Y}}_k\right)}{\sqrt{{\displaystyle \sum _{j=1}^N{\left(X_{ij}-{\overline{X}}_i\right)}^2}}\sqrt{{\displaystyle \sum _{j=1}^N{\left(Y_{jk}-{\overline{Y}}_k\right)}^2}}}$

with -1≤r _ik ≤+1 where $\overline{X}$_i is the mean over all X _ij 's for all j from 1..N, and $\overline{Y}$_k is the mean over all Y _jk 's all j from 1..N. In our context, X _ij and Y _jk are presence or absence of a COG within an organism or the percent response of an organism subject to adverse growth condition, respectively.

While the correlation measures the strength of association between an organism's genomic content and its phenotype, we also applied another method, exploiting the hypergeometric distribution function, to determine the significance of these associations

$p(i\ge m|N,M,n)={\displaystyle \sum _{i=m}^n\frac{\left(\begin{array}{l}M\\ i\end{array}\right)\left(\begin{array}{l}M-n\\ n-i\end{array}\right)}{\left(\begin{array}{l}N\\ n\end{array}\right)}}$

where N is the total number of species, n is the number of species that are positive in the laboratory result, M is the number of species that have the COGs family, and m is the number of species that have the COGs family and is also positive in the laboratory result, where a result ≤ 20% response is considered negative. So for a given gene found in M species, the hypergeometric function provides the probability by random chance that the gene is found in m species which contain the COG and are also positive in the laboratory test.

Assessment of predicted results

The following criteria were applied to the correlated data set. The intersection between a specific COG and a phenotype had to contain at least 3 organisms, and for any intersection, 30% of the microbes had to share the COG. The scores were adjusted using the standard Bonferroni error correction for multiple testing. Since the Bonferroni correction is one of the most conservative, it is likely that some biologically relevant associations were unnecessarily discarded. In this case α was set as less than equal to 0.01, therefore, any hypergeometric distribution score less than or equal to 0.0001 was deemed significant. Using these criteria, we set a 0.8 and a 0.9 correlation threshold to assess the significance of the COG-phenotype associations.

For the 0.8 correlation threshold, 290 total associations were obtained. We identified a subset of these data (154) that contain only COGs that have a known function. Of these 154 pairs, we performed detailed literature searches on 100 randomly selected pairs to confirm the validity of the positive COG-phenotype associations.

There were 74 associations found in the 0.9 correlation data set. Thirty-six of these pairs contained a COG of known function. Literature searches were performed on all 36 associations.

Website

All the GIDEON data used in this analysis is freely available from http://gersteinlab.org/proj/phenome .