EasyGene – a prokaryotic gene finder that ranks ORFs by statistical significance
 Thomas Schou Larsen^{1, 2}Email author and
 Anders Krogh^{1, 3}
https://doi.org/10.1186/14712105421
© Larsen and Krogh; licensee BioMed Central Ltd. 2003
Received: 25 November 2002
Accepted: 3 June 2003
Published: 3 June 2003
Abstract
Background
Contrary to other areas of sequence analysis, a measure of statistical significance of a putative gene has not been devised to help in discriminating real genes from the masses of random Open Reading Frames (ORFs) in prokaryotic genomes. Therefore, many genomes have too many short ORFs annotated as genes.
Results
In this paper, we present a new automated genefinding method, EasyGene, which estimates the statistical significance of a predicted gene. The gene finder is based on a hidden Markov model (HMM) that is automatically estimated for a new genome. Using extensions of similarities in SwissProt, a high quality training set of genes is automatically extracted from the genome and used to estimate the HMM. Putative genes are then scored with the HMM, and based on score and length of an ORF, the statistical significance is calculated. The measure of statistical significance for an ORF is the expected number of ORFs in one megabase of random sequence at the same significance level or better, where the random sequence has the same statistics as the genome in the sense of a third order Markov chain.
Conclusions
The result is a flexible gene finder whose overall performance matches or exceeds other methods. The entire pipeline of computer processing from the raw input of a genome or set of contigs to a list of putative genes with significance is automated, making it easy to apply EasyGene to newly sequenced organisms. EasyGene with pretrained models can be accessed at http://www.cbs.dtu.dk/services/EasyGene.
Keywords
Background
As of February 2003, 106 microbial genomes have been sequenced and made publicly available and the race is now on to mine genomes such as these for interesting and/or valuable genes and motifs. It has been estimated [1] that 60–80% of the genes in newly sequenced organisms have known homologues in other species. This percentage will grow as genomic annotations progress and perhaps there will be a time when virtually all genes can be found by homology matches to known proteins. That day, however, is not around the corner and even if it were, the occasional odd genes which would nevertheless escape detection by homology may very well be the truely novel and most wanted ones.
It is a common misconception that identification of genes in prokaryotes is almost trivial. Any random sequence, as well as noncoding regions in real genomes, contain a large number of open reading frames (ORFs). Most of these are too short to be possible protein coding genes, but in many genomes there are many 'random' ORFs longer than e.g. 100 amino acids, a cutoff that is often used for considering an ORF a real gene. The large number of short 'random' ORFs makes is difficult to discriminate real genes from random ORFs below a certain length, which depends on the genome and in particular its GC content. Therefore many genomes are overannotated [2]. In one genome, that of A. pernix [3], all ORFs longer than 100 amino acids are annotated as genes, but probably only around half the annotated genes are real [2]. The most severe problem today is to disciminate between short genes and random ORFs, and here the meaning of 'short' is quite organism dependent. One of the most important contributions of this paper is a way to deal with this problem by introducing a statistical significance for an ORF being a gene.
Computational gene finding exploits the statistical differences in codon usage between coding and noncoding regions of DNA [4–6]. The search for a mathematical frame work to efficiently capture these differences in codon usage led to Markov chain models and the GeneMark algorithm [7]. In order to facilitate the combination of various Markov chain scores, the application of Hidden Markov Models (HMMs) to gene finding was introduced in a gene finder for E. coli, Ecoparse [8]. These methods relied on a set of known genes for estimating parameters. More recently methods have been developed which start from a raw genome and automatically extract data for estimation. One of these, Glimmer [9], employs interpolated Markov models in order to use the maximum Markov chain order which can reliably be estimated for every oligomer. Another one, Orpheus [1], appeared the same year and calculates coding potentials of ORFs based on codon frequency of similarityderived genes. Most of these gene finders also extract ShineDalgarno sequences in order to improve prediction of start codons.
Due to their modular structure, HMMs are a suitable frame work for gene finding, and they are now used in GeneMark.hmm [10], GeneMarkS [11, 12] and FramebyFrame [13]. GeneMarkS uses a mixture of Markov chains of order 0, 1 and 2 in combination with features of already annotated genomes to bootstrap an initial estimation of coding statistics, which is then iteratively improved by the GeneMark.hmm2.1 algorithm. Gibbs sampling is also used to detect Ribosome Binding Sites (RBS). The FramebyFrame method was conceived to improve the accuracy of GeneMark.hmm and it employs standard Viterbi parsing of all six reading frames independently and subsequently combines these into a global parse.
In this paper, we describe a fully automated method for making an organism specific gene finder. It starts from a raw genome and searches for protein matches to get a good training set. Then an HMM with states for coding regions as well as RBS is estimated from the data set. This HMM is used to score all the ORFs in the genome and finally the score is converted to a measure of significance – R – which is the expected number of ORFs that would be predicted in one megabase of random DNA. The main advantage of this significance measure is that it takes the length distribution of random ORFs properly into account. The method is shown to match or exceed other gene finders currently available.
Methods
Automatic extraction of training set
 1.
Extract the maximal ORFs longer than 120 bases from the querygenome. For every stop codon, extract the region from the first downstream (in frame) start codon to the next downstream (in frame) stop codon.
 2.
Translate the ORFs to proteins and search for significant protein matches in SwissProt [14] using BLASTP without gaps and a threshold of 10^{5} [15]excluding proteins from the querygenome and proteins listed with one or more of the keywords putative, unknown, possible, hypothetical, probable, bacteriophage, transposon, insertion, reverse transcriptase.
 3.
For each ORF with at least one significant protein match, identify the most upstream position of these matches (to compensate for random matches, we actually remove 6 bases from the most upstream match before proceeding). If there is no alternative start codon between this position and the start of the ORF, put it in set A' of genes with certain starts. The remaining ORFs are put in set B' of certain genes with uncertain starts.
 4.
Reduce similarity in set A' by removing genes with similarity within the set. All pairs of genes are compared using BLASTN with a significance cutoff of 0.001. Two genes that match are called neighbours. Genes are removed starting with the one having the largest number of neighbours. This continues until no gene has neighbours left in the set [16]. The reduced set is henceforth referred to as A.
 5.
Unite set A' and B' into set T' and reduce similarity of T' to obtain set T .
 6.
Add 50 bases of upstream flank to genes of all sets and 10 bases downstream flank.
This procedure is a means to identify ORFs in a genome which are almost certain to be (protein coding) genes. The ORFs in set A and T make out reliable and balanced sets of positive examples which may be used to estimate the model parameters.
HMM architecture
We use a hidden Markov model (HMM) to model the gene structure. An HMM is a probabilistic model in which it is possible to model the various types of signals in a gene in one integrated model. For introductions to HMMs, see e.g. [17–19]. We model standard 'text book genes' with an unbroken open reading frame, i.e. genes with no programmed frame shifts, no sequencing errors nor any other special feature obstructing the reading frame.
The null model depicted on the left side of figure 2 has a third order state for capturing intergenic regions and a reverse codon model for modelling reverse genes ('shadows') with states of second order. Note that transitions are allowed directly from the null model to a start codon; this facilitates detection of genes inside operons which may not have a clear RBS.
We found that the inclusion of two more branches of internal codons improved performance. This is presumably because it allows the HMM to keep separate statistics for atypical genes, some of which may be horizontally transfered. Adding a fourth branch did not improve performance further, so we stopped at three lending some support to the hypothesis that there are essentially three classes of genes in prokaryotic genomes [22].
Model estimation
 1.
The emission probabilities of the background state are estimated from both strands of the complete genome.
 2.
The genes from start to stop codons are extracted from the training set and reverse complemented. The shadow model (consisting of three states) is estimated from these sequences. The parameters of this model are fixed.
 3.
The RBS, start and astart submodels (cf. figure 1) are estimated using set A exclusively since this set has reliable gene starts and therefore aligned upstream regions. Regions of 50 bases upstream from the start codon are extracted from all genes in set A. A null model (with fixed parameters) is included before the RBS model. The RBS model is initially primed with a high probability for the consensus GG (a dinucleotide common to most RBS), but the precise pattern is found by the estimation method. During estimation, a type of simulated annealing is used, where noise of a decreasing amount is added to the parameters in each iteration of the BaumWelch algorithm [23]. After training on set A, the parameters of the RBS, start and astart models are fixed.
 4.
The null model, RBS, start and astart models are now combined with the internal codons, bstop, stop and astop models to make up the complete model. The nonfixed parameters of this model are then trained on the (larger) set T .
The whole procedure can be completely automated. Note that no experimentally mapped RBSs are used for estimating the RBS model, the RBS is discovered during the estimation procedure.
When estimating the complete model (stage 4 above) we use labelled estimation [17, 20], where each base of the sequence is labelled as coding or noncoding. For the part of set T where the exact start is not known, we leave the part of the sequence from the most upstream start codon until the first significant protein match unlabelled. The weight with which each base in the unlabelled region contributes to the estimation of the parameters in the coding states is automatically determined during the iterative estimation procedure.
Decoding
By decoding we mean the process of finding an optimal parse of the DNA string into coding and noncoding regions.
The commonly used Viterbi decoding returns the most probable path of the sequence given the model [17], but this is not appropriate when the length is modelled by duplicating codon states, since this length modelling is realized only as a sum over many HMM paths. Therefore we use posterior decoding where one calculates, for each nucleotide i, the probability that it was emitted by a given state S. The calculation is done by adding the probabilities of all paths compatible with having state S emit nucleotide i [24, 17]. We use this to calculate the posterior probability of the first state of the start codon model, and thus obtain the probability that a gene starts at any position in the sequence. Given our assumption of perfect 'textbookgenes' with no errors or frame shifts, there is exactly one stop codon for each start, and thus the probability of a gene start is equal to the probability of the whole gene. The independent scoring of start codons makes it trivial to report several possible start codons for a gene in cases where there is no clear "winner".
Note that the HMM architecture in figure 1 is nonlooped – ie. it would find only one gene if we were using Viterbi decoding, which only gives the single most likely parse. This architecture however, is the correct one for scoring ORFs with posterior decoding. It has the further advantage that overlapping genes are easily handled since each gene start will be scored independently whether or not it overlaps other genes. In contrast, using a looped model and Viterbi decoding would not facilitate detection of overlapping genes unless the model contains explicit states for overlapping genes as described in e.g. [8].
The state posterior probability itself is not a useful score, because it is a probability of the whole sequence, not just a single gene, and it therefore depends on the length of the sequence it is part of. By dividing the posterior probability by the probability of the whole sequence (the genome) according to the null model, the contribution to the state posterior probability of the sequences flanking a gene will cancel and effectively make the ratio independent of the flanking sequences (except the parts very close to the gene), see the Appendix. The log of this ratio is called the logodds score, and that is our basic score for a gene.
Significance
As mentioned above, it is important to take into account the chance that a random ORF of the same length scores as high as a given gene. This is implicitly taken into account by our HMM because it models the length distribution of genes, but it turns out that one can calculate a significance score, which works slightly better (se results) and has a more intuitive interpretation.
The probability of finding highscoring ORFs in a random sequence is highly length dependent; the number of ORFs decays exponentially with the length so there will be a lot more short ORFs than long ones. For a random sequence of given length (e.g. 1 Mb) the expected number of ORFs of length l' can be written as
N(l') = exp (A  Bl'), (1)
where A and B are constants that can be found from linear regression of the log of the number of ORFs against the length. All lengths are measured in codons and we count both start and stop codon. For convenience we introduce the variable length l', which is the length of the ORF modelled by the looped codon submodels. A number of codons l_{0} are modelled explicitly in the beginning and end of the ORF (l_{0} = 4 in our model), so l' = l  l_{0}.
If the logodds score is denoted β, we show in the appendix that β' = β  (n  1) log(l'  n/2) is approximately normally distributed with a mean α_{μ} + γ_{μ} l' and variance α_{σ} + γ_{σ} l', which are both linear in l' (the α's and γ's are constants). Here n is the number of looped codon submodels (3 in our model). The coefficients of the linear mean and variance are estimated by linear regression on ORFs from random sequences. Then we define the standard score
which is normally distributed with mean 0 and variance 1.
For any given length l' and standard score Γ, one can now estimate the expected number C(l', Γ) of ORFs of the same length scoring higher than Γ in a fixed (long) random sequence using equation (1),
C(l', Γ) = exp(A  Bl') [1  Φ (Γ)], (3)
where Φ is the cumulative normal distribution. This number can be used directly to judge the significance of a gene predicted with length l and a standard score Γ. However, it would be preferable to know the total number of expected genes (of all lengths) predicted in a random sequence, rather than the expected number with a certain length. Therefore, instead of using C(l', Γ) to judge significance, we use the total number of expected predictions in a random sequence.
Suppose a gene of length l_{1} is predicted with a standard score Γ_{1}. Then the expected number of genes of that length predicted in a random sequence is C = C( , Γ_{1}). Now we want to calculate the total number of genes of all lengths predicted in a random sequence at the threshold C. For any length above a certain l_{ C }, the expected number of ORFs will fall below C due to the exponential decay of the length distribution. Therefore the total number of predicted genes (regardless of length) is roughly l_{ C }C + the sum of predicted genes above l_{ C }(a sum of a geometric series). We end up with the following expression for the total number of predicted genes in a random sequence:
The number 16 arises from the fact that the minimum ORF length considered is 20 codons corresponding to a variable length of 20  4 = 16. See Appendix for details. This is the expected number of genes predicted in a random sequence for a given value of C, and that number is the one we quote for each predicted gene. The constants A and B characterize the random sequence.
We have chosen to normalize R to a random sequence of 1 Mb rather than a random sequence the length of the genome, because then significance can be compared across genomes. The precise recipe we use for calculation of statistical significance is:

For the genome in question, generate 40 Mb of random sequence using the 3rd order background model (estimated from the genome). Extract all ORFs and estimate the parameters A and B by linear regression of the log of the number of ORFs against the length and normalize to 1 Mb. To avoid distortion of regression lines due to noisy statistics of long ORFs, restrict the variable ORF length used to lie below 70 codons. This makes the range of variable length 16–70 codons, which corresponds to a total length range of 20–74 codons.

Score all the ORFs in the random sequence with the model and calculate β' for all ORFs. For each length, calculate the average and variance and estimate the parameters α_{μ}, γ_{μ}, α_{σ}, and γ_{σ} by linear regression, again using ORFs in the range 20–74 codons.

To calculate the significance of an ORF in the genome, first calculate the standard score Γ from equation 2, then C from equation 3, and finally the significance value R using equation 4.
There are of course other possible choices of significance measure, but we believe that this is a simple and intuitively clear one and we have prefered it to the more traditional significance measure. By reporting the number R of expected false positives in one megabase of random sequence, it is easy for the enduser to estimate the number of false positives in a random sequence the length of the entire query genome – one simply has to multiply R with the size of the genome measured in megabases.
Using other gene finders
In order to benchmark EasyGene we compare it with some of the existing gene finders.
GeneMark 2.4, FramebyFrame, GeneMark.hmm 2.1 and GeneMark.hmm/S all belong to the GeneMark suit of programs and are accessible via the web interfaces listed in the references.
For GeneMark.hmm2.1, GeneMark.hmm/S and FramebyFrame the output is a coordinate listing (start and stop positions) of all predicted genes.
GeneMark2.4 outputs a list of stop codons and corresponding highscoring start codons. Each start/stop is listed with scores for coding potential and RBS. We collect all starts for a given stop and output the "Avg Prob" of the start with the highest RBS score. Whenever a threshold was needed for comparison purposes, we used 0.5 which is the default set on the web page.
Glimmer2.02 and Orpheus2 were installed locally. We changed the minimum ORF length predicted by these gene finders to 60 bp which seems to be the minimum used by the other gene finders. Orpheus and Glimmer provide two kinds of output: a verbose coordinate list of starts, stops and ORF scores and a simpler, postprocessed list of coordinates for ORFs regarded as genes.
In order to test their ORF scoring we had to parse the scored output. We had some difficulties interpreting the scored Orpheus output since some ORFs were scored several times with identical results (several identical "Start chosen"). In cases where multiple copies were found, we simply chose one of them and used the corresponding "Coding potential" (with the recommended threshold of 1) for further analysis.
For Glimmer2.02 the scored output was parsed simply by selecting the Gene Score attributed to every scored ORF and using the recommended threshold of 90.
Finally, we used RBSfinder [25] for an alternative postprocessing of Glimmer2.02 output. RBSfinder is designed to look for RBS sequences upstream of genes predicted by Glimmer2.02. If there are no RBS patterns in this region, RBSfinder searches for a start codon having a RBS pattern in the same reading frame upstream or downstream and relocates the start codon accordingly. The program may be iterated several times using revised predictions as new inputs. We found that running it twice was better than once but running it three times did not improve things further, so we chose to run it twice (with default options).
Results and Discussion
Figure 6 indicates that 3 branches is the best choice, while figure 7 suggests that the optimal order for these branches is 4. Note that the Yaxes of these figures have been zoomed in order to allow visual inspection of the performance differences. Note also that all architectures and orders in fact yield a relatively high true positive rate even at false positive rates below 0.02. Similar figures were also made for other organisms (not shown) and although the results were not always as clear as for E. coli, a choice of 3 branches of 4th order states as default models of the internal parts of genes works well for the organisms we tested.
True and false positive rates for selected Rvalues.
Rvalue  TP rate  FP rate 

0.1  0.971  0 
2  0.980  2.3e6 
10  0.984  3.0e5 
50  0.987  3.1e4 
150  0.991  8.7e4 
500  0.995  2.8e3 
10000  0.999  0.059 
Specificity, sensitivity and precision estimates for different gene finders in E. coli.
Data set  EasyGene  Glim  rbsGlim  Orpheus  Gm24  GmS  Gmhmm  Frame 

A'% found  98.4  98.9/98.9  98.9  98.0/95.3  91.5  97.2  98.1  97.0 
A'% exact  93.8  98.9/95.3  84.1  95.1/92.4  41.6  88.0  85.7  93.2 
B'% found  98.4  98.5/98.6  98.6  95.9/96.5  90.2  96.6  97.2  96.4 
T% found  98.1(98.0)  98.3/98.4  98.4  96.5/95.6  89.8  96.3  97.1  96.1 
Genome  4145  6827/5756  5756  9333/7543  3552  4064  4230  4064 
zero order  7  169/211  211  6761/5430  6  153  1459  0 
first order  7  545/723  723  6836/4804  13  241  830  0 
third order  1  2423/2694  2694  6582/4817  43  659  866  1 
shadows  0  19/21  21  22/9  1  0  2  0 
Sequence set A' consist of the 1136 genes with highconfidence starts extracted from the E. coli genome as described in Methods. The "% exact" row indicates the percentage of predictions where both start and stop codon are correctly predicted whereas the "% found" row indicates that only the stop codon is correct. Note that all genes in set A have the most upstream start (they are Longest Possible Open Reading Frames – LPORFs) by construction, and hence performance on this set favours gene finders which are biased towards LPORFs (such as Glimmer). Set B' consists of 1690 highconfidence genes with uncertain starts extracted as described in Methods. Due to the uncertainty of start codon placement, one cannot evaluate the exact start prediction performance for this set. The same is true for set T which is a similarity reduced union of A' and B' (2042 sequences).
The Genome row shows the number of genes predicted in the E. coli genome using default parameters and thresholds for the various gene finders. For Glimmer and Orpheus the minimum length of predicted genes was lowered to 60 in order to make their performance comparable to the others'. The next three rows show the number of false positives found in both strands of 1 Mb random sequences generated by zero, first and third order Markov chains estimated from the entire E. coli genome. The last row of the table shows the number of predictions wholly within the shadows of set A' – i.e. wholly within regions complementary to the genes in set A' where, ideally, no genes (or at least very few) should be predicted.
The test sets overlap with the training sets for EasyGene. Therefore, for set T the 10fold cross validation sensitivity is shown in parenthesis for EasyGene and it is seen to be reassuringly close to the noncrossvalidated sensitivity suggesting that EasyGene employs an appropriate model complexity and steers free of overfitting. Note also that the other gene finders have also been estimated from sets that overlap (or even contain) set T .
For Orpheus and Glimmer we show two numbers N1/N2 for each entry corresponding to before and after postprocessing. For Orpheus, N1 is the number of unique ORFs having a Coding Potential above the recommended minimum of 1 and N2 is the number of entries in the postprocessed orfnuc file. The postprocessing removes some overlaps but also appears to employ a less restrictive cutoff than the recommended 1. For Glimmer, N1 is the number of ORFs with Gene Scores greater than or equal to the recommended threshold of 90 and N2 is the number of entries in the postprocessed list of putative genes. The postprocessing elects ("votes") some ORFs as gene candidates despite a low score. On the other hand the postprocesing removes some samestrand overlaps in different reading frames so the N2 may be greater or less than N1 depending on the relative extent of overlaps and "voting".
It is always difficult to asses the specificity of a gene finder, because it is difficult to find genomic regions that are certain to contain no genes. We have therefore assessed specificity in three different ways. First, by counting the number of predicted genes in a genome. If this number is much higher than the number of annotated genes, it is likely that there are many false positive predictions, i.e. poor specificity. Our second test is based on random sequence. Clearly, a high number of predicted genes in a random sequence of bases indicates a poor specificity. However, it is probably not possible to find an exact quantitative correspondence between predictions in random sequences and real genomes. Also, it is not clear what sort of random sequence to use for such a test. By 0'th order we mean a sequence with bases generated randomly and independently with the base frequencies of the genome. Bases are quite correlated in DNA sequences, so we have also tested on sequences that are generated by Markov chains of orders 1 and 3. These Markov chains are estimated from the genome, so the sequences will have the same distribution of dinucleotides and 4nucleotides, respectively, as the genome. Finally, the third test is the number of genes predicted on the opposite strand of genes (shadows); these shadow regions should contain very few genes if any.
All gene finders except EasyGene, Frame and GeneMark2.4 predict a rather large number of false positives in random sequences, but for GeneMark.hmm and GeneMarkS we do not see large overprediction in the genome or in shadows. Evidently, Glimmer and Orpheus predict a lot more genes in the genome than the other gene finders, suggesting that these gene finders have very high false positive levels. This is supported by the high numbers of genes predicted in random sequences, and (to a much lesser extent) in shadows. Orpheus and Glimmer actually predict more genes pr. nucleotide in the third order random sequence than they do in the genome, suggesting that the coding potential calculated in these gene finders is far from optimal.
The HMM used by Frame assumes a minimum gene length of 69 bases which could make its false positive level seem somewhat better (lower) than it is, but there was no convenient way to lower the minimum length so we simply left it. It should also be noted that Frame relies on preexisting annotations for training and is therefore not a selftraining gene finder like Glimmer, Orpheus, GeneMarkS and Easygene.
The sensitivity of the gene finders is tested on sets of highconfidence genes. Glimmer has the highest sensitivity for all sets, but this is largely due to heavy overprediction. One ought always to bear in mind that it is not difficult to achieve high sensitivity if high levels of false predictions are tolerated at the same time – sensitivity is 100% if all ORFs are predicted as genes! Although there are some very close competitors, EasyGene comes out as the second best in sensitivity for all sets.
The exact prediction of start codons is tested on set A' and on an experimentally verified set. As mentioned above, set A is biased, because all genes of this set are LPORFs. Glimmer always predicts the most upstream start and consequently achieves a high performance on this set. When Glimmer's output is postprocessed by RBSfinder the performance drops considerably.
Sensitivity and precision estimates for experimentally verified E. coli genes.
Data set  EasyGene  Glim  rbsGlim  Orpheus  Gm24  GmS  Gmhmm  Frame 

LiC% found  100  100/100  100  97.7/91.7  97.7  100  100  100 
LiC% exact  94.0  100/97.0  88.7  96.2/90.2  49.6  94.0  90.2  98.5 
LiD% found  100  100/100  100  96.8/98.4  100  100  100  100 
LiD% exact  96.8  0/1.6  75.8  51.5/51.6  67.7  95.2  80.6  87.0 
Specificity, sensitivity and precision estimates for M. tuberculosis, H. pylori J99 and B. subtilis.
M. tuberculosis  

Data set  EasyGene  GmS  Frame 
A'% found  96.7  97.2  96.0 
A'% exact  89.1  80.9  87.9 
B'% found  96.8  97.1  96.3 
T% found  96.9(96.6)  97.3  96.4 
Genome  3749  3983  4341 
zero order  0    8 
first order  3    2 
third order  2    12 
shadows  1  0  1 
H. pylori J99  
Data set  EasyGene  GmS  Frame 
A'% found  99.2  99.2  99.2 
A'% exact  97.5  95.7  96.7 
B'% found  100  99.6  98.9 
T% found  99.7(98.8)  99.5  99.1 
Genome  1491  1518  1479 
zero order  2  1479  2 
first order  1  336  2 
third order  0  403  0 
shadows  2  0  0 
B. subtilis  
Data set  EasyGene  GmS  Frame 
A'% found  99.3  98.1  98.8 
A'% exacts  94.8  94.1  93.3 
B'% found  99.2  99.0  98.2 
T% found  99.3(99.2)  99.0  98.4 
Genome  4083  4221  4006 
zero order  1  675  0 
first order  2  457  0 
third order  1  813  2 
shadows  0  0  0 
For M. tuberculosis GeneMarkS and EasyGene are comparable, although GeneMarkS seems to overpredict slightly (assuming that the 3918 annotated genes are close to being correct), and Easygene might underpredict. We believe that Frame predicts too many genes in this organism and at the same time it has lower sensitivity than the two others, suggesting a worse performance overall. This indicates that Frame is not very robust with respect to high GC content.
The small dataset from H. pylori might give a slight overfitting in EasyGene, where there is one percent difference between crossvalidated results and noncrossvalidated. For this organism the three gene finders seem to have very similar performances. Finally, for B. subtilis EasyGene comes closest to the number of annotated genes and have higher sensitivity than the other methods.
Conclusions
The emerging overall picture is that the sensitivity of EasyGene tends to be comparable to GeneMarkS and higher than Frame. With regards to specificity, EasyGene and Frame tend to be comparable and both higher than GeneMarkS. Hence, EasyGene comes out with the best combined sensitivity/specificity performance. When it comes to exact starts, EasyGene also generally performs best.
Glimmer and Orpheus have good sensitivities but at the cost of very low specificities in this comparison. We have lowered the ORF length cutoff from their default values in these methods to make the results comparable. This may be unfair, but since the challenge is to find the short genes, we believe that any gene finder should be able to score them.
At present it is not possible to automatically find all genes in a prokaryotic genome. We believe the aim of a gene finding system is to help expert annotators as much as possible, and we consider the statistical significance of a gene an important help in classifying the predictions into almost certain genes and borderline genes needing more attention. Contrary to most other gene finders discussed here, it is up to the user which significance cutoff to use. EasyGene also predicts suboptimal start codons if need be, so it will be easy to see if e.g. two alternative starts have almost the same score.
A shortcoming of the significance value as calculated here is that long ORFs score well simply on account of their length, since very long ORFs occur rarely in random sequences. For this reason, EasyGene also provides a logodds score in the output which may be held up against the Rvalue for ORFs longer than say 500 bp. Genes of this length ought to have high logodds values. If they do not – i.e. if they score below 0 – then they are probably not real genes despite their length (sometimes very long noncoding ORFs occur in regions of repetitive DNA).
If the amount of available genomic DNA is very small (as it may be in partially sequenced genomes) the 3 branches of 4th order coding models may have too many parameters to be reliably estimated. In such cases, one could reduce the parameter space simply by using fewer branches and/or lower orders. More generally, one could develop a method for automatic finetuning of HMM architecture for every new organism. Other lines of future research could focus on modelling of genes with errors and frame shifts.
Finally, it may be noted that a prototype of EasyGene has already been used in the annotation of S. typhi [31].
Appendix: The length dependent score distribution
The logodds score distribution has several components. The probability of a sequence z containing and ORF given a model M may be written as
where c_{1}, . . . , c_{ l }denotes the codons in the ORF and flanks denotes all the rest of the sequence. P (lM) is the length distribution of the HMM. In logarithmic form we have
Similarly, the denominator of the logodds score, log P (zN ), may be rewritten in exactly the same way, so the entire logodds score becomes
where
We will now look at the distribution of each of these terms.
The null model consists of a state with a loop and three reverse codon states with a loop. For a long sequence one of the loops will usually dominate the probability, so the length distribution is well approximated by a geometric distribution
where k_{1} is a constant greater than zero.
The length distribution from the looped codon states is a negative binomial [17],
where p is loop probability and n is the number of looped codon submodels (in casu, n = 3). Some of the ORF (start and end) is modelled by nonlooped states. The number of nucleotides modelled in the beginning and end of the ORF is denoted l_{0} and the variable length is denoted l', so the total ORF length is l = l_{0} + l'. Using that log (n  1) log(l'  n/2)  log(n1)! for l' >>n, we obtain
The next term, Q, in equation 6 is a sum of random variables since we are considering random sequences. The sum Q will therefore (according to the central limit theorem) converge to a normal distribution N (μl, σ ), for large l.
The submodels flanking the gene model are identical to the null model. Therefore p(flanksM) and p(flanksN) will almost cancel in the last term of equation 6, except from the contribution from the RBS model and the states after the stop codon. This contribution will be independent of the ORF length. Since it is again a sum of random terms it is well approximated by a normal distribution.
Apart from the nonlinear term from the negative binomial we have now shown that all terms in equation 6 are constant or scale linearly with l. Therefore
is normally distributed for fixed length with an average and variance that are linear functions of l:
average = α_{μ} + γ_{μ}l' (11)
The parameters of these linear functions can readily be estimated from random sequences by doing linear regression between the variable length and the mean and variance of logodds scores as shown in the two upper plots of figure 9. (Note that one can switch from the variable length l' to the total length l in the formulas by changing the constants.)
Finally it is convenient to introduce the standard score Γ, defined in equation 2, which is normally distributed with average 0 and variation 1.
Since random ORF lengths are geometrically distributed (cf. equation 1), the expected number of ORFs of length l scoring more than Γ in a sequence is
C(l', Γ) = exp (A  Bl') [1  Φ(Γ)], (13)
where the exponential term is the expected number of ORFs of variable length l' and Φ is the cumulative normal distribution. Thus, for a given score we can calculate the number of expected predictions in a random sequence.
For lengths l ≥ l_{ C }this expectation is always less than C (due to the exponential factor). So if we require that no more than C predictions be made at any length, the total number of predictions in a random genome is
where l_{ s }is the shortest ORF length considered (we use 20). l_{ C }is given by
R = {(A  log C)/B  (l_{ s } l_{0})}C + C/(1  exp(B)). (16)
Web sites used
FramebyFrame:
http://opal.biology.gatech.edu/GeneMark/fbf.cgi
GeneMark.hmm 2.1:
http://opal.biology.gatech.edu/GeneMark/gmhmm2_prok.cgi
GeneMark.hmm 2.1 using GeneMarkS models:
http://opal.biology.gatech.edu/GeneMark/gmhmm2_genemarks.cgi
GeneMark 2.4:
Declarations
Acknowledgements
We wish to thank G. Kolesov for kindly providing the Orpheus package and the people behind http://www.tigr.org/software/glimmer for providing the Glimmer package. This work was supported by a grant from the Danish National Research Foundation.
Authors’ Affiliations
References
 Frishman D, Mironov A, Mewes HW, Gelfand M: Combining diverse evidence for gene recognition in completely sequenced bacterial genomes. Nucleic Acids Research 1998, 26(12):2941–2947. 10.1093/nar/26.12.2941PubMed CentralView ArticlePubMedGoogle Scholar
 Skovgaard M, Jensen LJ, Brunak S, Ussery D, Krogh A: On the total number of genes and their length distribution in complete microbial genomes. Trends in Genetics 2001, 17(8):425–428. 10.1016/S01689525(01)023721View ArticlePubMedGoogle Scholar
 Kawarabayasi Y, et al.: Complete genome sequence of an aerobic hyperthermophilic crenarchaeon. Aeropyrum pernix K1. DNA Res 1999, 6: 83–101.View ArticlePubMedGoogle Scholar
 Fickett J: Recognition of protein coding regions in DNA sequences. Nucleic Acids Research 1982, 17: 5303–5318.View ArticleGoogle Scholar
 Gribskov M, Devereux J, Burgess R: The codon preference plot: Graphic analysis of protein coding sequences and prediction of gene expression. Nucleic Acids Research 1984, 12: 539–549.PubMed CentralView ArticlePubMedGoogle Scholar
 Staden R: Measurements of the effects that coding for a protein has on a dna sequence and their use in finding genes. Nucleic Acids Research 1984, 12: 551–567.PubMed CentralView ArticlePubMedGoogle Scholar
 Borodovsky M, McIninch J: GENMARK: Parallel gene recognition for both DNA strands. Computers and Chemistry 1993, 17(2):123–133. 10.1016/00978485(93)85004VView ArticleGoogle Scholar
 Krogh A, Mian IS, Haussler D: A hidden Markov model that finds genes in E. coli DNA. Nucleic Acids Research 1994, 22: 4768–4778.PubMed CentralView ArticlePubMedGoogle Scholar
 Salzberg SL, Delcher AL, Kasif S, White O: Microbial gene identification using interpolated Markov models. Nucleic Acids Research 1998, 26(2):544–548. 10.1093/nar/26.2.544PubMed CentralView ArticlePubMedGoogle Scholar
 Lukashin AV, Borodovsky M: GeneMark.hmm: new solutions for gene finding. Nucleic Acids Research 1998, 26(4):1107–1115. 10.1093/nar/26.4.1107PubMed CentralView ArticlePubMedGoogle Scholar
 Besemer J, Lomsadze A, Borodovsky M: GeneMarkS: a selftraining method for prediction of gene starts in microbial genomes. implications for finding sequence motifs in regulatory regions. Nucleic Acids Research 2001, 29(12):2607–2618. 10.1093/nar/29.12.2607PubMed CentralView ArticlePubMedGoogle Scholar
 Besemer J, Borodovsky M: Heuristic approach to deriving models for gene finding. Nucleic Acids Research 1999, 27(19):3911–3920. 10.1093/nar/27.19.3911PubMed CentralView ArticlePubMedGoogle Scholar
 Shmatkov AM, Melikyan AA, Chernousko FL, Borodovsky M: Finding prokaryotic genes by the 'framebyframe' algorithm: targeting gene starts and overlapping genes. Bioinformatics 1999, 15(11):874–886. 10.1093/bioinformatics/15.11.874View ArticlePubMedGoogle Scholar
 Bairoch A, Apweiler R: The SWISSPROT protein sequence database and its supplement TrEMBL in 2000. Nucleic Acids Research 2000, 28(1):45–48. 10.1093/nar/28.1.45PubMed CentralView ArticlePubMedGoogle Scholar
 Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller W, Lipman DJ: Gapped BLAST and PSIBLAST: a new generation of protein database search programs. Nucleic Acids Research 1997, 25(17):3389–3402. 10.1093/nar/25.17.3389PubMed CentralView ArticlePubMedGoogle Scholar
 Hobohm U, Scharf M, Schneider R, Sander C: Selection of representative protein data sets. Protein Science 1992, 1: 409–417.PubMed CentralView ArticlePubMedGoogle Scholar
 Durbin RM, Eddy SR, Krogh A, Mitchison G: Biological Sequence Analysis. Cambridge University Press, Cambridge, UK 1998.View ArticleGoogle Scholar
 Krogh A: An introduction to hidden Markov models for biological sequences. In Computational Methods in Molecular Biology (Edited by: Salzberg SL, Searls DB, Kasif S). Elsevier, Amsterdam 1998, 4: 45–63.View ArticleGoogle Scholar
 Rabiner LR: A tutorial on hidden Markov models and selected applications in speech recognition. Proc IEEE 1989, 77(2):257–286. 10.1109/5.18626View ArticleGoogle Scholar
 Krogh A: Two methods for improving performance of a HMM and their application for gene finding. In Proceedings of the Fifth International Conference on Intelligent Systems for Molecular Biology (Edited by: Gaasterland T, Karp P, Karplus K, Ouzounis C, Sander C, Valencia A). Menlo Park, CA AAAI Press 1997, 179–186.Google Scholar
 Yada T, Totoki Y, Takagi T, Nakai K: A novel bacterial genefinding system with improved accuracy in locating start codons. DNA Research 2001, 8(3):97–106.View ArticlePubMedGoogle Scholar
 Medigue C, Rouxel T, Vigier P, Henaut A, Danchin A: Evidence for horizontal gene transfer in Escherichia coli speciation. Journal of Molecular Biology 1991, 222(4):851–856.View ArticlePubMedGoogle Scholar
 Hughey R, Krogh A: Hidden Markov models for sequence analysis: extension and analysis of the basic method. CABIOS 1996, 12: 95–107.PubMedGoogle Scholar
 Stormo GD, Haussler D: Optimally parsing a sequence into different classes based on multiple types of evidence. In Proc of Second Int Conf on Intelligent Systems for Molecular Biology 1994, 369–375.Google Scholar
 Suzek BE, Ermolaeva MD, Schreiber M, Salzberg SL: A probabilistic method for identifying start codons in bacterial genomes. Bioinformatics 2001, 17(12):1123–1130. 10.1093/bioinformatics/17.12.1123View ArticlePubMedGoogle Scholar
 Link AJ, Robison K, Church GM: Comparing the predicted and observed properties of proteins encoded in the genome of escherichia coli k12. Electrophoresis 1997, 18: 1259–1313.View ArticlePubMedGoogle Scholar
 Hannenhalli SS, Hayes WS, Hatzigeorgiou AG, Fickett JW: Bacterial start site prediction. Nucleic Acids Research 1999, 27(17):3577–3582. 10.1093/nar/27.17.3577PubMed CentralView ArticlePubMedGoogle Scholar
 Cole ST, et al.: Deciphering the biology of Mycobacterium tuberculosis from the complete genome sequence. Nature 1998, 393: 537–544. 10.1038/31159View ArticlePubMedGoogle Scholar
 RA Alm, et al.: Genomicsequence comparison of two unrelated isolates of the human gastric pathogen helicobacter pylori. Nature 1999, 397: 176–180. 10.1038/16495View ArticleGoogle Scholar
 Kunst F, et al.: The complete genome sequence of the grampositive bacterium bacillus subtilis. Nature 1997, 390: 249–256. 10.1038/36786View ArticlePubMedGoogle Scholar
 Parkhill J, et al.: Complete genome sequence of a multiple drug resistant salmonella enterica serovar typhi ct18. Nature 2001, 413(6858):848–852. 10.1038/35101607View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article: verbatim copying and redistribution of this article are permitted in all media for any purpose, provided this notice is preserved along with the article's original URL.