Endless modifications of existing antibiotics might lead to the appearance of cross-resistance; therefore there is a need for parallel efforts of developing new types of antimicrobials. The preceding step is often finding a new drug target for these drugs. However, given the wealth of information provided by genome sequencing, mass spectrometry and microarray experiments, the selection of a potential drug target for pathogenic species and their relatives is not a trivial task. The repertoire of approaches starts with simple similarity searches, during which sequences of bacterial proteins are compared to known drug targets and human proteins 1 . Other methods are more extensive and implement analyses of metabolic and interaction networks 2 - approaches that, to our knowledge, were first tested with human protein drug targets 3 . There have been attempts to use machine learning methods to mine the substantial amount of data that can be found and derived for bacterial proteins. These approaches may focus on proteins as targets, providing lists of features (sequence length, mostly beta or alpha secondary structure, cytoplasmic/membrane bound, enzyme/non-enzyme, etc.) for a typical drug target 4 . Alternatively, they can focus on protein-protein interactions 5 . Finally, given a small molecule, one can dock it to the set of protein structures described in 6 , which aids in the identification of new, potential drug targets for known compounds.

In general, all essential proteins of a given organism constitute potential drug targets 7 8 . The most prominent, essential proteins, such as the ribosomal proteins, are already approved drug targets in bacteria 9 . Their importance for the cell results in their evolutionary conservation 10 . For over 30 years, it has been commonly thought that essential genes are likely subject to stronger negative (purifying) selection (with a less frequent occurrence of mildly deleterious substitutions) 11 compared to nonessential genes. However, the prokaryotic and eukaryotic kingdoms seem to differ in this respect. No statistical difference in the relative rate of evolution between essential and nonessential genes was evident for the mouse, if immune genes were excluded 12 . Analysis of the yeast genome sheds light on eukaryotic species, by explaining why it was difficult to see a statistically significant difference in that case 13 (they observed a significantly higher rate of evolution of nonessential proteins when they compared essential proteins with the “most dispensable” half of nonessential proteins). In bacteria, stronger negative selection on essential genes was shown by Jordan 14 in the case of Escherichia coli K12. By applying an orthology-based essentiality transfer from Escherichia coli, the authors also predicted that this should be true for pathogenic species in the Neisseria and Helicobacter genera. This finding has been used for prediction of essential genes on its own or in conjunction with other methods.

No doubt we can learn a lot about choosing good drug targets from nature. Antibiotics are mainly originating from natural fungal and bacterial products 15 . Microorganisms have been using them for millions of years to combat (successfully) competing organisms. This impressive finding has been confirmed recently using paleontological data 16 . From the evolutionary point of view it may suggest that good drug targets are evolutionary constrained and are subject to purifying selection, which makes them less susceptible to random development of resistance. The efforts to validate this hypothesis and to find its application in drug design workflows comprise the aim of this study.

Analysis of evolutionary rates to identify putative drug targets has been already suggested by Searls 17 , but no comprehensive study has been published so far. Moreover, two methods were proposed to identify evolutionary constrained residues in drug targets. Durand and co-workers 18 assessed purifying selection on individual sites in Plasmodium falciparum drug targets using the dN/dS ratio. The method is called “evolutionary patterning” (EP). A second method called “evolutionary tracing” (ET) 19 , was proposed by Lichtarge in 1996. The Lichtarge method is based solely on evolutionary conservation. In both methods the key assumption is that a good drug should bind to the slowly evolving protein pocket. This expectation is based on intuition that at such sites development of drug resistance will be less probable. Both methods do not provide the overall picture of the evolutionary rates of genes of pathogenic species as they focus on individual sites, not the whole genes.

Generally speaking the key question we wanted to answer in this study was whether proteins which are targeted by antibiotics tend to evolve slowly. For this purpose we analyzed relative rate of evolution of genes from seven bacterial pathogens and from E. coli. We used polymorphism analysis, i.e. pN/pS ratio (which represents an appropriate measure of purifying selection in the case of comparison of inter-species diversity) and reproduced the analysis with dN/dS ratio (which is better for comparing sequences derived from different species) 20 . In both cases the rate of evolution of known drug targets, was not only significantly lower than the genome average but was also significantly lower than that for the essential genes, suggesting a higher selective force acting on a wide spectrum of drug targets. This finding suggests that calculation of evolutionary rate can aid in scoring during the process of drug target selection and can provide additional insights into whether a particular protein might or might not be an attractive drug target. As such, it complements EP/ET approaches.

Antibiotics are mainly natural products used by micro-organisms against other micro-organisms. They seem to be relatively evolution proof, i.e. resistance is sufficiently rare and it is still beneficial for microorganism to use antibiotics against competing microorganisms.

We have shown that in most of the analyzed pathogenic genomes potential drug targets have statistically significant higher negative selection than essential genes or the genome average. Our explanation for this phenomenon is that such proteins are evolutionary constrained, i.e. they are overall highly sensitive to perturbations, which could correspond to relatively infrequent point mutations (including those leading to resistance).

One may ask the question whether our observation is not the result of the fact that bacteria have already been subject to considerable “drug” pressure - either by medical usage of compounds 43 or by more ancient and long term exposures to the natural products on which the antibiotics are based 44 . The observed purifying selection may, in this case, be the result of selective sweeps brought about by the drugs. If it would be the case more sensitive variants would be removed due to selection caused by the drug. The dN/dS analysis suggests that this possibility should be excluded. This test compares relative rate of evolution observed in the comparison of two closely related species. In contrast to the pN/pS test which takes in account existing allelic diversity, dN/dS takes into account only fixed mutations (with frequency = 1 in the population). Therefore, it is much more robust to the observed selective sweeps caused by antibiotics.

In some cases in our study we observed exceptions to the described general pattern. For example for Staphylococcus aureus and Haemophilus influenzae we did not observe a statistical difference in average pN/pS values between essential genes and potential drug targets (although drug targets evolve much slower than other genes). It is likely that for many drug targets, directed positive selection has led to intrinsic resistance; many Staphylococcus aureus strains are known to be resistant to vancomycin (VRSA) as well as methicillin (MRSA). Similarly, many Haemophilus influenzae strains have intrinsic resistance to beta-lactam drugs. It makes the observable differences in evolutionary rate between drug targets and other analyzed groups (essential genes and all genes) being lower than they in fact are.

Drug targets also have a higher negative selection when assessed by dN/dS ratio (omega). We consider these results supplementary as omega analysis has certain drawbacks (e.g. sensitiveness to recombination, limited range of dS for which dN/dS ratio is considered to be reliably estimated or non-linear dependency on time). Nevertheless, dN/dS analysis confirms the results obtained by the pN/pS approach.

When thinking about developing drugs against a certain target, one must also consider issues such as resistance mechanism (efflux pumps, other resistance proteins), drug target accessibility or host-related factors. Considering whether a gene is essential (even in a broader context than is commonly considered 45 ) or conserved is often not sufficient. The assessment of evolutionary rate (e.g. by pN/pS values) helps substantially in the evaluation of potential drug targets. The resulting targets have an evolutionary history suggesting that they are less likely to randomly develop resistance via point mutations. And while it seems that the only cases one could find are the “obvious” ones, we show that this approach identified metal ion binding genes and succinate dehydrogenases - neither of which corresponds to well-studied wide-spectrum drug targets. Our results show that pN/pS analyses are an attractive addition to drug target prediction pipelines.

On the other hand one should be aware of the biases in our method. First, we used only whole genomes to limit the cases where low quality alignments will substantially affect pN/pS estimation. Because of that the eight chosen species are among the most common infective bacteria (they were among the first sequenced species). This, of course, is advantageous to address concerns like MRSA and VRSA and other antibiotic resistant pathogens, but then there is no evidence that the conclusions of this study apply to the less common human pathogens, to veterinary pathogens, or to other bacteria which could be antibiotic targeted. Second, we needed to limit the analysis to only eight species as the approach undertaken requires data on essential genes. This raises a concern whether Gram-negative species dominating in this study led to biased results. And finally, predicting evolutionary rate with pN/pS is limited to alignments of closely related sub-species. The last issue can be avoided by performing dN/dS analysis, but as we mentioned above dN/dS methodology has its own drawbacks.

In this study we showed that good drug targets evolve slowly and that the rate of evolution is a better predictor of drugability than essentiality. This to some extent explains why known antibiotics (usually being of microbial origin) have been efficiently targeting other microorganisms for millions of years of evolution 16 .

Our study also shows that evolutionary rate can be used to score and find potential drug targets. Generally our approach can be considered a useful complementation to EP (Evolutionary Patterning) and ET (Evolutionary Tracing) approaches. Both those methods can be useful in designing a drug that targets a specific site and has a known mechanism of operation. Our approach can be considered an attractive solution in the preceding step, i.e. finding the targets which could be analyzed in detail by ET or EP.

The authors declare that they have no competing interests.

AG performed the analysis. AG, SK, PS and PZ participated in the design of the study. AG and PS drafted the manuscript. SK and PZ revised the manuscript. All authors read and approved the final manuscript.