One of the longstanding questions in mammalian genome evolution has been the origin of GC-isochors, which are long (> 100 kb) chromosome segments characterized by a high degree of uniformity in GC-composition levels [1, 2]. Several theories including selectionism [3], neutralism [4], thermodynamic stability [5], and GC-Biased Gene Conversion (gBGC) [6] have been proposed to explain the origin of isochors, but have not been conclusive. The gBGC hypothesis was initially formulated by Holmquist [7] and Eyre-Walker [8, 9], and has been elaborated upon since then [10, 11]. gBGC is proposed to be a consequence of special cases of meiotic recombinations that involve the formation of heteroduplexes [11, 12]. A heteroduplex is created when a short single-stranded DNA segment of one of the parental chromosomes forms a double stranded structure with its complementary homologous strand from the equivalent chromosome of another parent. The presence of SNPs within a heteroduplex results in mismatched base pairs that are resolved by the molecular machinery of the DNA mismatch repair (MMR) pathway [13]. gBGC hypothesizes a bias in the repair, in which mismatched nucleotide pairs are resolved in favor of G-C pairs [10]. This would imply that mismatches involving non-Watson-Crick base pairs such as A-G, T-C, A-C or T-G would be preferentially repaired to yield C-G, G-C, G-C and C-G Watson-Crick base pairs respectively. In 2013, Lesecque and co-authors experimentally confirmed the existence of such bias in yeast, yet it appeared to be very small (50.6% vs 49.4% in AT - > GC base pair changes vs. GC - > AT ones) [14].

One important consequence of gBGC is the increased overall GC-content at recombination hotspots [15], where heteroduplexes are more frequent than elsewhere. There are both supporting and opposing evidences for gBGC. The supporting propositions posit that gBGC explains the evolution of non-randomness in GC-compositions within mammalian genomes [11], the rapid fixation of AT- > GC mutations [11, 16] and increased GC content of recombining DNA in mammals and yeast [17]. Evidence of gBGC has also been reported in yeasts [18], arabidopsis [19], and honeybee genomes [20]. On the other hand, conflicts with the gBGC hypothesis have also been published. For example, analysis of GC/AT and AT/GC substitutions in the human Fetuin-A gene ruled out gBGC as one of the causal factors [21]. Population genomic analysis of Drosophila melanogaster revealed no evidence for gBGC [22] and, in another instance, non-allelic gene conversion processes in Drosophila and primate genomes also negated the contribution of gBGC towards organism diversity [23]. A negative correlation between substitution and recombination rates in the chicken genome was also reported, inconsistent with the gBGC model [24]. Even though gBGC with respect to humans has been reported in recombination hotspots and rapidly-evolving regions of the human genome [25, 26], a quantitative picture of gBGC inside any region of the human genome was lacking until recently [11]. In the last 2 years, new publications reported directly observed cases of gBGC events in several large families [27, 28]. These authors demonstrated a strong bias in AT- > GC over GC- > AT base pair changes during non-crossover gene conversions in humans.

In the modern time when whole-genome sequencing is a routine and thousands of human genomes are available, is it possible to detect and quantify gBGC events in a large scale? One of the fundamental principles in genome organization is that neighboring SNPs are linked into haplotypes. A gene conversion event results in replacement of an allele inherited from one parent by the complementary allele inherited from another parent. About half of gBGC episodes occur during crossover (CO) miotic recombinations while another half during non-crossover (NCO) recombinations. In the latter case of NCO, the gBGC allele replacement occurs within the same parental haplotype producing only single nucleotide change inside this haplotype. We call this specific process a “SNP flip-over” throughout this paper. In other words, SNP flip-over is an allele replacement event for a single SNP in the middle of evolutionarily conserved haplotype. Such SNP flip-over may be directly detected by comparing the genomes of mother, father, and offspring, when haplotypes of all three are available. However, since the New Generation Sequencing technique is dependent on assembling millions of short reads, haplotypes are computationally deduced in a so-called “phasing” procedure. It is a statistical prediction that generates a vast number of phasing errors [29–31]. Phasing problems make the detection of de novo gBGC events difficult. To overcome this problem Williams with coauthors and Halldorsson with coauthors studied three-generation pedigrees with multiple family members [27, 28]. Such direct detection of gBGC is of ultimate importance for the validation of this phenomenon yet it provides limited statistics.

On the other hand, Glemin and co-authors used analysis of derived allele frequency from the 1000 Genomes data to quantify gBGC in human [32]. We chose another approach for detection of relatively recent SNP flip-overs, due to gBGC that happened hundreds or thousands years ago. We computationally analyzed only common haplotypes built from frequent SNPs that remained unchanged in different populations for thousands of years. Then, we looked for a very rare haplotype that is practically identical to one of the common haplotypes except with one allele replacement (flip-over) at one of the polymorphic sites in the middle of the haplotype. We called this specific type of rare haplotype as “Acceptor” haplotype (See Fig. 1). We searched 1092 sequenced genomes for people that have the acceptor haplotype from one parent and its nearly identical common haplotype counterpart inherited from another parent (as illustrated in Fig. 1b). This requirement for parental haplotype organization is important for avoiding possible errors due to low sequencing coverage in the 1000 Genomes dataset. Indeed, under such conditions all polymorphic sites within a haplotype under analysis are homozygous except the one “Acceptor” site representing putative gBGC conversion event. This homozygosity requirement eliminates possible phasing errors.

In this paper, we performed bioinformatics investigation of SNPs in the “1000 Genomes” database and observed 223,955 putative cases of gBGC events. We report here a quantitative assessment of the aftermath of gBGC in.

Genotype datasets for all the human chromosomes of the 1092 human genomes were downloaded from the 1000 Genomes ftp site (ftp://ftp-trace.ncbi.nih.gov/1000genomes/ftp/release/20110521/) [33] as Variant Call Format (VCF) files version 4.1 [34]. This database contains a total of 38.2 M SNPs, 3.9 M short indels and 14 K deletions for all the human chromosomes that have been used in this study. Information about parental haplotypes has been taken directly from Phase 1 of 1000 Genomes Project, since its genomic sequences were entirely “phased”. We considered only bi-allelic variants for simplicity and because we had sufficient statistics in our datasets. All multi-allelic variants were skipped. Allele frequency for every SNP was obtained from the “AF=” field inside column 8 of the 1000 Genome VCF files. We did not take into account differences in population frequencies in our paper because this frequency variation is not relevant at all for finding SNP flip-over events, which is the major focus in this study.

All haplotypes of 1092 individuals, putative cases of gene conversion and local GC content of putative gene conversion sites were obtained with our pipeline of five Perl programs (HaploFind.pl, GeneConversionFind.pl, BGC_Calculator.pl, LocalGC_calculator.pl, AT_vs_GC.pl, RandomGC_Calculator.pl). Detailed description and scripts of all our Perl programs, their instruction manuals, the command lines for execution of programs, and examples of output files can be found in the Additional file 1.

Haplotypes for every chromosomal segment were computationally constructed from 50 adjacent frequent genetic variants (MFA > 25%). Since more than 90% of these frequent genetic variants are presented by SNPs, we call these variants as SNPs for simplicity throughout the text. However, for putative BGC events having “flip-over” alleles inside common haplotypes we considered only SNPs (all indels were rejected from consideration).

Our computer characterization of common haplotypes has been elaborated in the predecessor project [35]. In that research we tried different parameters for the threshold for frequent SNPs (Minor allele frequency (MAF) 10%, 20, 25, or 30%) and the number of frequent SNPs in the haplotypes (30, 50, or 70). In the present paper we chose the default parameters of the previous research (MAF = 25%, number of SNPs in the haplotype = 50), which allow identifying an optimal amount of common haplotypes. The number of common haplotypes and the linkage disequilibrium between SNP within a common haplotype varies from one chromosomal segment to another as described in Dutta et al., [35].

The exact of base pair changes from Common haplotype to Acceptor haplotype have been characterized by the Perl script GeneConversionFind.pl.

Where N is total number of gene conversion events (AT ➔ GC and GC ➔ AT), p is the proportion of AT ➔ GC events and (1 – p) is the proportion of GC ➔ AT events.

All our programs are freely available from our website (http://bpg.utoledo.edu/~afedorov/lab/BGC.html) [37]. The entire dataset of all haplotypes for each 56,328 chromosomal segments generated by our programs is also available from this web site (this dataset is too big to place it in Additional file 1).

What is the average number of SNP flip-overs in a human gamete? The estimation of this number is essential because SNP flip-overs change SNP-haplotypes and linkage disequilibrium between SNPs. This effect should be taken into consideration in various programs used for deciphering phenotypes from SNP patterns because noticeable SNP flip-over process constantly modifies these patterns and reduces linkage disequilibrium between neighboring SNPs. SNP flip-over occurs during NCO meiotic recombination events when one allele is replaced by its counterpart allele, while the neighboring SNPs in the haplotype remain the same. NCOs are up to 15 times more frequent than COs [41–45]. A recent study estimated 228 NCO events on average per generation in humans [27]. At the same time, average length of NCO heteroduplex tracts are much shorter than CO tracts with average NCO tract length of 75 bp according to several recent studies [27, 44, 46]. Therefore, we estimate 228 × 75 bp = 17.1 kb of total NCO heteroduplex length per gamete. On the other hand, the number of heterozygous sites for Europeans and Asian individuals in the 1000 genomes dataset are about 2.3 million, and about 3.3 million for African individuals [34]. Taking these groups together, on average there is about one heterozygous site per 1.2 kb in the human genome. Considering all the above, in human meiosis, about 14.2 mismatches (17.1 kb/1.2 kb) should be formed within all NCO heteroduplexes of a gamete. During repair, only half of these 14.2 mismatches should resolve into SNP flip-overs, while in the other half of cases, MMR should restore the original alleles within original haplotypes. This leaves 7.1 SNP flip-overs per gamete. They represent up to a quarter of all new mutations in a gamete (there are from 20 to 50 novel mutations in a gamete according to different estimations [47–49]). Hence, SNP flip-over has a substantial impact on nucleotide changes and should be considered in any SNP dynamics analyses.

The second important question we address is the number of AT ➔ GC versus GC ➔ AT base-pair conversions per human gamete due to gBGC. To estimate this number, we should consider both CO and NCO cases since both results in base pair conversions. The estimated sex-averaged number of COs per generation is ~ 30 [50]. We will use the average CO heteroduplex tract length of 600 bp for humans, which is consistent with several current studies [46, 51]. So, we estimate in total 30 × 600 bp = 18 kb of CO heteroduplex tracts length per gamete. Thus, number of mismatches formed in all CO heteroduplexes during human meiosis is about 15 (18 kb / 1.2 kb). In the previous paragraph we already calculated that the average number of mismatches in NCO heteroduplexes is 14.2 per gamete. Since, half of mismatches should be resolved in base-pair changes, the total number of base-pair changes due to both NCO and CO will be, on average, 14.6 events per gamete. According to our calculations (Table 1), 15.4% of SNP flip-overs do not cause AT/GC base pair changes, 45.7% create AT- > GC base pair replacements, while 38.9% are responsible for the reverse GC- > AT replacements. Thus, on average, 14.6 base-pair change events should generate 2.2 cases with no GC/AT changes (i.e., GC < −>CG and AT<− > TA), 6.7 cases with AT- > GC base pair conversions, and 5.7 cases of GC- > AT conversions. In sum, every human gamete should generate an excess of 1 AT- > GC base pair changes due to gBGC episodes. This number is twice the estimates for yeast genomes [14]. At the same time, our assessment is lower than that of Williams and coauthors who evaluated that 68% heterozygous AT/GC SNPs transmit GC alleles [27]. The overabundance of 1 AT- > GC base pair changes per gamete, if occurring over millions of years in mammals, may yield significant biases in GC-compositions along chromosomes. However, this effect should be evaluated only in combination with an influx of de novo mutations, which is roughly 20–50 mutations per gamete [47–49]. Thus, overabundance of 1 AT- > GC base changes per gamete due to gBGC represents only 2–5% of all new mutations in the gamete genome and might be over-shadowed by other mutational processes. Distribution of novel mutations is uneven along the genome and depends on the local nucleotide composition at the site of mutations. Our laboratory reported a strong fixation bias favoring AT - > GC mutations in GC-rich regions in humans and the opposite fixation bias favoring GC - > AT mutations in AT-rich regions and other fixation biases (e.g. Pu - > Py in pyrimidine rich regions) [40]. Therefore, estimation of the total effect of mutations and conversions on the genomic GC composition is very intricate and still awaits thoughtful modeling.

During the process of meiotic non-crossover recombination, a human gamete acquires about 7 SNP flip-over events, in which one allele is replaced by its complementary allele while the neighboring SNPs in the haplotype remain the same. On an average, GC-Biased Gene Conversion increments the GC-content by substitution of one AT pair by one GC pair in every haploid human genome. Happening over millions of years of evolution, this smallest bias may be a noticeable force in changing the nucleotide composition landscape along chromosomes.