Published isochore datasets show remarkable diversity. In the following we will use the human genome for comparisons of different isochore assignments if not stated otherwise. The number of isochore segments found in the human genome ranges from about 1,200 for GC-Profile to up to more than 76,000 for BASIO. As a consequence, the resulting isochores show very different length distributions. Isochores discovered by least-squares segmentation are the longest at an average of 2,459 kb, whereas BASIO and IsoFinder segments are the shortest at an average of 40 and 72 kb, respectively (Figure 1). It can be seen that IsoFinder and BASIO are clearly in a different league compared to GC-Profile and least-squares in terms of the number and average length of isochores. This divergence results from different criteria used by the four tested methods to determine the beginning and end of the segments, and the window lengths of 10 and 100 kb used by BASIO and least-squares, respectively. As explained in Materials and methods, a difficult challenge in GC-content-based partitioning of complex eukaryotic genomes is to find a set of parameters suitable for coping with the significantly different levels of GC fluctuations in the GC-rich and GC-poor regions.

Using the GC level of each isochore, we evaluated the GC difference (delta GC) between adjacent segments and found that the delta GC distributions of the compared methods are significantly different. The BASIO and the least-squares data show the smallest GC jumps while the GC-Profile and IsoFinder methods produce the broadest distribution and the greatest delta GC values on average (Figure 2). One explanation for this may be that short isochores are more likely to model local GC outliers, which results in higher delta GC differences between adjacent segments, on average.

We further assessed the differences between the segmentation methods based on the entropy distance between them. Lower entropy distance values indicate a better agreement between two isochore maps. As shown in Table 1, the results of the least-squares and BASIO approaches are the most dissimilar as measured by this criterion. It is noteworthy that the positions of about 25% of the borders of the least-squares map are identical to those of the BASIO segmentation. This exact border coincidence is an exception, however; in most of the cases segment borders are shifted by between 10 kb and 100 kb for the methods. No borders are shifted by more than 1 Mb with regard to the BASIO borders (Additional data file 1).

Despite the striking differences between the isochore assignments in terms of segment borders and isochore length, a strong agreement exists with regard to the amount of equally classified DNA and genes. As shown in Table 2, all four original methods assign about 66% of the human genome to the same isochore families. The isochore families are described in detail in the Materials and methods. Furthermore, the four methods locate around two-thirds of all genes in isochores of the same family (Table 3). On average, the consensus in attributing genes to the same isochore between each individual method and the three other methods is between 60.1% (IsoFinder) and 62.4% (least-squares).

The breakdown of the genome into the five isochore families is very similar for all the methods. On average, 22 ± 2.5% (standard deviation) of the complete human DNA is found in the L1 isochore. The most dominant isochore family is L2, with 34 ± 2.7% of the DNA, followed by the H1 family with 23 ± 1.5%. The remaining 15% of the genome is distributed between the H2 and H3 families, with 11.4 ± 0.2% and 3 ± 1.1% of the DNA, respectively. The low deviation values among the methods indicate a good overall agreement between all the isochore maps.

Significant similarities between the DNA and gene classifications produced by the different computational methods render a consensus isochore assignment feasible. As outlined in the Materials and methods, the consensus assignment assumes the isochore family that is predicted by the majority of methods at each genomic position. This simple consensus approach results in 31,176 distinct isochores in the human genome, with an average isochore length of 99 kb (Figure 1). The median and average delta GC differences between neighboring isochores are 4.1 and 4.6, respectively (Figure 2). With regard to the number, length and delta GC values of isochores, the consensus assignment shows a reasonable balance between the observed extreme values of the individual methods. The amount of ambiguous DNA, that is, the nucleotides that could not be classified by the majority approach, is less than 0.2%. Our interactive online isochore browser (Figure 3) allows for a visual comparison between the individual isochore assignment methods and the consensus isochore map.

Due to the lack of large-scale experimental data on isochore location in the human genome, we are evaluating whole-genome isochore assignments using indirect evidence by considering independent biological properties known to be associated with GC content variation. One such property is gene density (the number of genes per Mb) which is known, to vary significantly between different isochore families of the human genome 52627, from very high in H3 to very low in L1. This observation was first made experimentally and subsequently confirmed by genome sequencing; for a review of possible causes, see 24272829. A biologically meaningful genome segmentation would thus be expected to display a strong correlation with gene density.

We compared the different isochore maps with respect to the degree of correlation between genome segmentation and gene density. As an example, Figure 4a shows a comparison between GC-Profile and the consensus method. Both methods display a clear dependence on the isochore classification of genomic regions, with gene density varying over a broad range between 5 (for both GC-Profile and the consensus map in the L1 isochore) and 73 or 92 (for GC-Profile and the consensus map, respectively, in the H3 isochore). The consensus assignment thus conforms better to the intuitive isochore-gene density model in that it displays higher gene density in the H3 isochore (Figure 4a). Therefore, the consensus isochore assignment provides a stronger signal in terms of gene density-isochore correlation than the GC-Profile segmentation.

The strength of the correlation between two variables can be estimated in a more rigorous way based on the slope of their respective linear regression lines, as shown in Figure 4b. The greater the slope of the consensus regression line the stronger the association between the resulting segmentation and gene density compared to GC-Profile. As seen in Table 4, the slope of the consensus isochore map is steeper than that of all other methods, signifying that the consensus approach is the most valid one with respect to this particular biological feature.

In addition to our genome-wide analysis of gene density, we carefully analyzed currently available direct experimental evidence pertinent to isochore properties (Table 5). For each of the five computational methods (IsoFinder, GC-Profile, BASIO, least-squares, and the consensus approach) we investigated whether or not they meet the respective criteria. The first two tests took advantage of the recent experiments of Schmegner et al. 30. In their work, they showed that the human MN1 gene (residing in a GC-rich isochore) is replicated several hours earlier (during the S phase of the cell cycle) than the neighboring gene PITPNB from a GC-poor isochore. Furthermore, a second isochore border within the human KIAA1043 gene was described and experimentally verified. As seen in Table 5, the first border between MN1 and PITPNB was correctly recognized by all methods except for the least-squares approach. The second border in the KIAA1043 gene was not detected by the least-squares or the GC-Profile assignments. We are aware that these failures may be overcome by further tuning of these methods, although this will give rise to a host of new questions. However, all isochore borders are correctly found by the consensus approach. In a further test, we checked the detection of the well known isochore border between the genes encoding the human MHC class II and class III region 17. This border is correctly found by all methods. This is not surprising as all methods were evaluated against the available body of experimental evidence at the time of their publication and fine-tuned by their respective authors.

Finally, we evaluated the isochore length distributions. Early experiments that applied fragmentations at various scales 23 as well as theoretical studies 18 suggest a typical isochore length significantly longer than the average size of 72 and 40 kb predicted by IsoFinder and BASIO, respectively, in the human genome. GC-Profile and least-squares meet these isochore length requirements. However, none of the individual methods - except for the consensus method - results in an isochore map that shows an isochore length distribution similar to that annotated by the Bernardi group for an outdated human genome assembly 18. As summarized in Table 5, the consensus approach appears to be more robust in that it meets all experimentally verified criteria, while all other methods fail in one or more tests. Furthermore, the quality of the consensus assignments is bound to further improve as more complementary isochore prediction methods are incorporated.

Most genes completely reside within a single isochore stretch (Additional data file 2). A comparison of random segmentations that have comparable block lengths shows that more genes are wholly located within an isochore segment than would be expected by chance. This is especially pronounced in isochore segmentations with segments of relatively short average length, such as those determined using IsoFinder and BASIO, and underlines the utility of isochore information for gene prediction. This observation may be related to the structure of chromatin 31 or chromosome break-prone regions 32. We also found that most genes are classified into the same isochore families by the different methods. As a consequence, the isochore assignment confidence, as defined in Materials and methods, is very good for most genes and hardly any genes are classified with low confidence (Figure 5). One further observation is that most genes are found in regions with integer confidence values. This can be explained by the fact that genes typically reside completely within a single isochore stretch, irrespective of the applied method. For example, if a gene is completely covered by an isochore stretch in all isochore predictions, then the confidence value for this gene will always be two, three or four, depending on the number of methods that agree in their classification. In contrast, non-integer confidence values indicate regions that show a certain agreement for parts of the gene only, usually because an isochore border is located within a given gene. Overall, 99.8% of all genes are assigned to the same isochore families by at least two methods. This provides a sound basis for using isochore classification of genes in experimental studies such as expression analysis.

Overall, the confidence of the isochore assignment in the human genome is higher in GC-poor regions (Figure 6). The confidence decreases in GC-richer regions and reaches a minimum at GC content values around 55-58%. This may be explained by the increasing GC fluctuations in GC-richer regions 33. Elevated confidence levels corresponding to the lowest and highest GC levels may be explained by simple statistical reasons. For example, the GC-richest regions are most likely to be classified into one out of two isochore families: the GC-richest H3 family or the less GC-rich H2 family. By contrast, a segment with an intermediate GC content may fall into one of three isochore families (for example, H2, H1 or L1). Given this limited event space, the likelihood of observing an agreement between the methods for the GC-richest and GC-poorest regions will be higher. The isochore confidence is least near isochore borders in all isochore maps (Figure 7). It quickly grows with distance from the borders and reaches saturation at a distance of approximately 0.2 Mb from the border. This empirical observation can be useful for defining a 'safe distance' threshold in practical applications of isochore information, allowing the estimation of the isochore classification reliability at any region of interest even if no consensus or confidence information is to hand.

We calculated isochore assignments and evaluated their confidence for 20 completely sequenced vertebrate genomes using GC-Profile, IsoFinder, least-squares and BASIO as well as our consensus method (Tables 6 and 7). The amount of DNA that could not be classified by majority vote into one of the five isochore families in our consensus maps for any of these 20 genomes was very small, less than 1% on average. The overall isochore assignment confidence is generally very high, with 2.6 methods agreeing on average. The entropy distance between the consensus map and the segmentations of all four individual methods indicates to which isochore segmentation the consensus map is most similar. This large-scale comparison shows that there is neither a single method clearly closest to the consensus, nor a simple dependency of a method's performance on the overall GC-richness of the genomes.

We furthermore present in Table 7 the amount of DNA that is found in each of the isochore families for all genomes. As expected, the overall GC content of a genome influences the amount of DNA in the different isochore families in that the genomes that have, on average, higher GC content are supposed to have more DNA in GC-richer isochores. However, a simple correlation could not be found. For example, in the dog genome, 5% of the DNA is in H3 isochores, whereas in the platypus genome only 1% is in the H3 isochores. The opposite would have been expected as the platypus genome has a high overall GC content (46%) in comparison to the much lower GC content (41%) of the dog genome.

We have created an online database, IsoBase, where all data described in this study are freely accessible. Our website enables the user to evaluate statistical distributions of isochore properties, and compare isochore assignments within and between organisms and methods. Multiple qualitative and quantitative properties of isochore maps can be interactively explored. Confidence values of each segment are displayed for each consensus isochore map. Tables 6 and 7 show an overview of genomes included in our database and their isochore properties.

For convenience, we provide two search interfaces at our IsoBase website 25. The first search feature allows the genomic positions and the isochore families of genes to be looked up by free text searches and by multiple identifier types. Currently, genes can be looked up by RefSeq identifiers, UniProt/SwissProt accessions, Ensembl IDs, gene and protein names, as well as by their descriptions, and SwissProt keywords. The second search option allows retrieval of available isochore information for a list of genomic positions in one step. All isochore assignments and the corresponding confidence information can be visualized online and downloaded as tab-delimited data files. In addition, we provide UCSC custom annotation tracks of the consensus isochore assignments for all genomes. All UCSC tracks can be downloaded from our web site. Furthermore, the isochore tracks are integrated into the UCSC view automatically by using the links to the UCSC genome browser at our web site 25.

We refer to the isochore nomenclature as it was first described based on ultra-centrifugation experiments 26. Bernardi and colleagues 18 defined the isochores according to their GC content. There are three isochore types with high GC content, H3 (>53%), H2 (46-53%), and H1 (41-46%), and two types with low GC content, L1 (<37%) and L2 (37-41%). In Additional data file 3 we present an analysis of the amount of genomic DNA versus segments' GC content (by 1% bins) and confirm that distinct isochore families can be observed throughout the genomes analyzed in this study. The Bernardi group 18 calculated the GC content of 100 kb long, non-overlapping sequence windows and then merged the windows if the difference in their GC content was below 1-2%. However, no hard threshold was used, and in many cases subjective decisions were made as to whether or not to merge windows, making the Constantini method as described in the original publication hardly fully automatable. In particular, this circumstance makes it impossible to consider the Constantini data for our comparison of isochore assignment methods, which is based on a more recent version of the human genome than the one used in the original publication.

In this work isochores were predicted by four methods for automatic genome segmentation: GC-Profile 2223, BASIO 21, IsoFinder 20, and least-squares optimal segmentation 1934. Briefly, GC-Profile is a windowless method that recursively partitions the input sequence into two subsequences, left and right, based on the quadratic divergence between statistical measures (such as genome order indices, a²+c²+g²+t², where a, c, g, and t are occurrences of individual bases) reflecting base composition. IsoFinder moves a sliding pointer along the input DNA sequence and finds a position that maximizes the GC difference between its left and right portions according to t-Student statistics. Then both portions are split into non-overlapping 300 kb windows, and for each individual window the GC content is computed. If the mean values of the window GC content on the left and the right of the pointer position are significantly different, this position becomes the cutting point and the input sequence is divided into two subsequences. Both GC-Profile and IsoFinder proceed from left to right and may produce different results if the direction is inverted. BASIO calculates Bayesian marginal likelihood for sequence segments and, for reasonably short DNA contigs, attempts to find a global maximum of the overall likelihood over all possible configurations of segment borders using a Viterbi-like dynamic programming algorithm. For large DNA sequences, such as complete chromosomes, BASIO relies on an approximate split-and-merge procedure to find an optimal segmentation. We applied the BASIO method using the default border insertion penalty 3 and 10 kb sequence blocks as initial input. Finally, the least-squares method calculates GC content (values logarithmized as in 19) in non-overlapping 100 kb windows (default window size as in 19) and then derives optimal segmentation of the resulting array of real values, which yields the minimal sum of squares of the Euclidian distance between each value and its segment average. However, the least-squares algorithm requires the user to provide the expected number of output segments as a parameter. As an estimate of this number for the least-squares method we utilized the minimum number of isochores produced by the three other methods - GC-Profile, BASIO, and IsoFinder. This approach makes over-fragmentation unlikely and provides a lower limit for the actual number of isochores. All methods are clearly distinct in terms of their methodology; a review of fundamental statistics in segmentation approaches is given in 35. Additionally we show in Additional data file 3 that all methods make a complementary contribution to the consensus maps throughout all genomes.

Methods that rely on any information beyond the raw nucleotide sequence for isochore prediction were not considered in this study. For example, the Markovian approach of Melodelima et al. 36 incorporates information about known biological features such as genes and their properties to create hidden Markov models. By contrast, all the methods in this study are solely based on the GC content and, therefore, can be used even in the absence of reliable gene models, for example, in a newly sequenced genome.

We used the human genome as a test case for comparing isochore assignments made by the different methods. The latest human genome assembly hg18 (build 36) was obtained from the UCSC genome browser 37. Further vertebrate genomes were downloaded from UCSC, Ensembl 38, and the Broad Institute 3940. Assembly parts marked as random and short scaffold parts were not considered. The 'UCSC known genes' models 41 were used for computing gene density, defined as the number of genes per million nucleotides (Mb). To determine the gene density in individual isochores, we counted the number of genes that start in each isochore family and divided it by the total amount of genomic DNA classified into the respective isochore family. For the regression analysis the isochore family labels were translated into their ordinal value: from 1 for the L1 family to 5 for the H3 family. Gene density values were logarithmized (natural logarithm) as they grow polynominally with increasing isochore family number. Statistical tests were performed using PROMPT 42.

In this study we are measuring the distance between two segmentations P and Q by the 'entropy distance' as introduced and discussed by Mielikäinen et al. 43 and Haiminen et al. 44, respectively. Briefly, the entropy H of a segmentation P with k segments can be defined as:

H ( P ) = − ∑ i = 1 k Pr ⁡ ( p i ) log ⁡ Pr ⁡ ( p i )