Among the 895 complete prokaryotic genomes available from the Kyoto Encyclopedia of Genes and Genomes (KEGG) (http://www.genome.jp/kegg/, July 2009 release), 829 genomes are accompanied with characterizations of temperature range conditions in NCBI’s Microbial Genome Project Database (http://www.ncbi.nlm.nih.gov/genomes/lproks.cgi). The temperature range of a prokaryotic organism is characterized by one of the five terms – cryophilic, psychrophilic, mesophilic, thermophilic and hyperthermophilic; however, the organisms are unevenly distributed under these five temperature range conditions, with the majority falling into the category mesophilic (Table 1). To maximally take advantage of the provided information regarding organisms’ temperature range conditions as well as to obtain statistically significant results, we excluded the cryophilic and psychrophilic categories from further analysis, merged 57 thermophilic and 36 hyperthermophilic organisms into the same category (labeled as thermo/hyperthermo-philic), and then focused on the 722 mesophilic versus 93 thermo/hyperthermo-philic organisms.

Among the 2,798,133 genes of the 815 organisms being studied, 1,210,908 (~43.3%) genes are assigned with KEGG Orthology (KO) IDs, and a total number of 6,026 KO groups are being covered. We considered genes with the same KO IDs as orthologous, so that they can be used to estimate the distribution of the GC content level surrounding each gene (corresponding to a unique KO ID) in various genomes. For each gene, we obtained the GC content levels of both the coding and surrounding non-coding regions, where the non-coding region starts from the end of the previous gene till the start of the next gene with the coding region of the current gene being excluded. As the coding and non-coding regions were treated independently, we used genomic region to refer to either of them in the rest of the paper.

From the statistical point of view, the significance of the correlation relationships between the temperature range condition and the GC content levels of the selected genomic regions can be measured by the p-values derived from the KS tests. On the other hand, if the temperature range condition of an organism is predictable based on the GC content levels in the selected genomic regions, then the correlation relationships between the two traits can be further justified.

For the prediction test, each organism was represented by a feature vector consisting of the GC content levels of the selected genomic regions. A cross validation procedure [34] was applied to estimate the prediction accuracy. That is, the collection of the mesophilic genomes and the collection of the thermo/hyperthermo-philic genomes were each randomly partitioned into two portions. The portions with 10% of the genomes from both collections were used for testing to estimate the classification accuracy, while the other portions with 90% of the genomes were used for building a SVM-based classifier. This partition-training-and-testing procedure was repeated for 500 times to obtain the average prediction accuracy. The estimate for the prediction accuracy was benchmarked against the prediction accuracy rendered by random guessing. To facilitate such comparisons, we enforced that for each cross validation experiment the same number of organisms (93 vs. 93) for both temperature range conditions were used so that the accuracy rate rendered by random guessing is 50%. Note that the 500 repeats of the cross validation procedure could allow us to sample a variety of combinations of mesophilic organisms so that the estimate of the prediction accuracy was unbiased to particular sub-collections of mesophilic organisms.

To test the generalizability of the identified correlation relationships, we also applied the SVM-based classifier to predicting the temperature range conditions of 17 mesophilic and 17 thermo/hyperthermo-philic in-progress genomes in NCBI’s Microbial Genome Project Database (see Table S-2 of the supplementary material), as well as 29 mesophilic and 29 thermo/hyperthermo-philic metagenomes in the Integrated Microbial Genomes (IMG/M) system [35] (see Table S-4 of the supplementary material). The prediction accuracy for these in-progress genomes and metagenomes was averaged cross the 500 SVM classifiers, each of which was derived from the training phase of the above-mentioned cross validation procedure.

A collection of 413 genomic regions (including the coding and non-coding regions of 80 genes, the coding regions of 197 genes, and the non-coding regions of 56 genes) were consistently detected in all the 200 bootstrap KS tests and were included into {g_complete}. To find out the biological implications underlying these 413 genomic regions, we performed enrichment analysis to determine which KEGG pathways are over-represented by the pertinent genes of these genomic regions. The enrichment factor of a KEGG pathway is defined as the ratio of the percentage of the genes involved in this pathway among the genes in {g_complete} to the percentage of the genes involved in the same pathway among all the genes being considered. And, a KEGG pathway was considered to be enriched if its enrichment factor is greater than 1. Here we summarize the KEGG pathways with the 10 largest enrichment factors in Table 2.

It can been seem from Table 2 that one of the most enriched KEGG pathways is cell motility. It has been reported that the temperature affects bacterial movements in both chemotaxis (ko02030) and flagellar assembly (ko02040). For instance, flagella synthesis in Escherichia coli, Proteus sp. and Salmonella sp. are all inhibited at higher incubation temperature [36]. And, as for chemotaxis behavior, most motors of a motile bacterial cell spin exclusively clockwise at very low temperatures so that the cell tumbles more frequently; and, when the environment temperature is incremented, the cell usually show some increase in average translational velocity [37,38]. Note that styrene degradation (ko00643), ubiquinone and menaquinone biosynthesis (ko00130), and streptomycin biosynthesis (ko00521) are also among the top 10 enriched KEGG pathways, suggesting that these pathways are also subject to thermal influence. In accordance with these computation-based findings are the following biological experiment-based findings that are documented in various publications: (i) An increase of the temperature from 32°C to 40°C effectively decreases the styrene degradation rate of Rhodococcus pyridinovorans PYJ-1 to ~66% of the optimum value [39]. (ii) Ubiquinone-8 formation in Escherichia coliK-12 strain AB 2847 was greatly affected by temperature variations – the rate of converting 2-octaprenyl phenol to ubiquinone-8 reached the maximum at 32°C, while virtually no reaction at all occurred at a temperature of 0°C [40]. And, (iii) Production of streptomycin in Streptomyces griseus ATCC 12475 was suppressed at elevated growth temperature, and even failed at 34°C or above [41]. These findings demonstrated that our computation-based analysis on the correlation between genomic and ecological traits may guide experimental investigations on the mechanisms of prokaryotic organisms adapting to environmental changes.

When based on the GC content levels of the genomic regions in {g_complete}, the prediction accuracy, averaged over the 500 cross-validation experiments, was 92.45% for the complete genomes, and was 92.05% for the in-progress genomes. Note that both of the accuracy rates were significantly higher than the 50% accuracy yielded by random guessing.

There were four genomic regions in {g_complete} consistently selected during all the KS tests on the partial lists of organisms with particular phylogenetic/oxygen requirement/habitat/salinity categories being excluded. Table 3 summarizes the mean and standard deviation of the GC content level in each of these four genomic regions for mesophilic as well as thermo/hyperthermo-philic organisms. Observe that the GC content levels in these four regions tend to be higher in mesophilic than thermo/hyperthermo-philic organisms, which confirms the hypothesis that prokaryotic organisms living in high temperature range conditions do not necessarily rely on high GC contents to cope with their environment [23,24].

These four genomic regions correspond to the coding regions of K01251 (adenosylhomocysteinase), K03724 (DNA repair and recombination proteins), K07588 (LAO/AO transport system kinase), and K09122 (hypothetical protein), and are interpreted as temperature-correlated but irrelevant to the distribution in various oxygen requirement, salinity, habitat and phylogenetic groups. Some of these computation-based findings can be further supported by experiment-based findings. For instance, K01251 reflects a common adaptation mechanism at high temperatures, because all the residues forming the network of aromatic and hydrophobic contacts and the residues potentially involved in cofactor binding are fairly well conserved among the hyperthermophilic but not in the mesophilic organisms [42]. Hyperthermophilic organisms may require DNA damage repair (K03724) to be unusually effective to cope with the destabilization of the DNA at high temperatures [43]. And, micro-array experiments on the Escherichia coli strains (record GDS1848) that had been entrained in the high temperature conditions (41.5°C degrees) also showed that the expressions of K03724 (lhr) and K07588 (argK) were significantly different than those of the wild strain that lives at 37°C degrees [44,45].

The phylogenetic trees of these four genes were built by using ClustalW2 [46] and Phylip [47], and are shown in the supplementary material. Based on these trees, we calculated the evolutionary distances between organisms of the same temperature range (intra-temperature) and the evolutionary distances between organisms of different temperature range conditions (inter-temperature). On average, the inter-temperature distances are 1.12, 2.29, 1.54, and 2.13; and, the intra-temperature distances are 0.79, 1.80, 1.16 and 1.48 for K01251, K03724, K07588, K0912, respectively. Therefore, in terms of the evolutionary trajectories of these four genes, organisms of the same temperature range conditions are closer to each other than they are to organisms of different temperature range conditions. When comparing the number of neighboring species with the same temperature range conditions in the phylogenetic trees built from the four genes against that based on the 16S rRNA genes (Table 4), we did not observe statistically significant differences, indicating that phylogenetically close species tend to inhabit the same temperature range conditions.

When based on these four genomic regions, the prediction accuracy, averaged on the 500 cross-validation experiments, was 84.52% for the complete genomes (84.46% for mesophiles, 85.09% for thermo/hyperthermo-philes), 84.09% for the in-progress genomes (83.05% for mesophiles, 85.13% for thermo/hyperthermo-philes), and 82.70% for the metagenomes (79.31% for mesophiles, 86.21% for thermo/hyperthermo-philes), respectively, all of which were much higher than the 50% accuracy rendered by random guessing. It should be noted that with less than 1% of the elements in {g_complete} as the features, the prediction accuracy only degraded 8.6% when compared to that rendered by the entire {g_complete}, suggesting that these four genomic regions are predominantly correlated with the temperature range condition.

When the KS tests were conducted on the mesophilic vs. thermo/hyperthermo-philic organisms based on the temperature range annotations in the GOLD database, eight genomic regions were consistently selected out of the tests on the complete list of organisms and on the partial lists of organisms with particular phylogenetic/oxygen requirement/habitat/salinity categories being excluded. These genomic regions include the coding regions of K01251 (adenosylhomocysteinase), K03724 (DNA repair and recombination proteins), K07588 (LAO/AO transport system kinase), K08289 (purine metabolism), and K09122 (hypothetical protein), and the non-coding regions of K00261 (glutamate dehydrogenase), K03809 (Trp repressor binding protein), and K07008 (currently unclassified). Note that the coding regions of K01251, K03724, K07588 and K09122 were the genomic regions consistently selected when the KS tests were based on the temperature range annotations in the NCBI’s Microbial Genome Project Database. That is, despite the 11.51% inconsistency in the temperature range annotations in the NCBI and GOLD databases, the four genomic regions were always identified as to possess GC content levels that are correlated with the temperature range conditions, suggesting their robustness to the possible annotation errors in the databases.

There were six genomes whose temperature range conditions were consistently mis-classified (in at least 80% of the cross validation experiments that they were used for testing), whether the prediction was based on the entire {g_complete} or just the four genomic regions. The optimal growth temperature and the temperature range conditions of these six genomes are summarized in Table S-3 of the supplementary material. This result may prompt us to re-check the current temperature range condition annotations of the NCBI’s Microbial Genome Project Database. For example, we did find evidence to support our prediction for Deinococcus geothermalis DSM 11300 and did observe inconsistent annotations in the NCBI database for the temperature range conditions of Methanococcus aeolicus Nankai-3 and Exiguobacterium sp.AT1b.Deinococcus geothermalis DSM 11300 is labeled as mesophilic in the NCBI’s Microbial Genome Project Database, but is actually thermophilic with an optimum growth temperature of about 45°C to 50°C [48]. Methanococcus aeolicus Nankai-3 grows at temperatures between 20-55°C [49], and Exiguobacterium sp. AT1b grows at temperatures between 15-55°C [50]. The growth temperatures of both Methanococcus aeolicus Nankai-3 and Exiguobacterium sp. AT1b cross the boundaries of NCBI’s definition for mesophilic and thermophilic organisms and overlap more with the mesophilic category. However, Methanococcus aeolicus Nankai-3 is labeled as mesophilic while Exiguobacterium sp. AT1b is labeled as thermophilic in the NCBI’s Microbial Genome Project Database. These examples suggest that the correlation relationships between certain genomic and ecological traits of prokaryotic genomes can potentially facilitate “reverse ecology” [51], i.e., to fill/refine the environmental annotations of prokaryotic organisms based on their genomic features.

Supplementary materials for the conditions of different environmental factors, identified genomic regions, in-progress and meta-genomes, and consistently misclassified genomes are available at: http://users.ece.gatech.edu/~hzheng7/TempRanGC.pdf.