Statistical analysis and simulations that account for simultaneous effects of positive, negative, and no crossover interference in multilocus recombination data Authors: Shaul Sapielkin, Zeev Frenkel, Eyal Privman, Abraham B Korol Published: 2025 ABSTRACT AbstractCrossover interference (COI) is a widespread feature of homologous meiotic recombination. It can be quantified by the classical coefficient of coincidence (CoC), but this characteristic is highly variable and specific to the pair of chromosomal intervals considered. Several models have been proposed to characterize COI at the chromosomal level. In the gamma model, the strength of interference is characterized by a shape parameter ν. In contrast, the gamma-sprinkled two-pathway model (GS) accounts for both interference-dependent and independent crossover (CO) events by fitting a mixture of gamma distributions with v > 1 and v = 1 with proportions 1-p and p. In reality, COI may vary along the chromosome, resulting in а poor fit of the employed model to the analyzed data. Additional inconsistency can be caused by the common neglect of possible negative crossover interference (NCOI) in the model, although the phenomenon was previously reported for several organisms. We present an extension of the GS model to account for possible NCOI and provide suitable software for estimating the parameters of the extended GS model. On real chromosome data from a conifer (larch, Larix principis-rupprechtii), sugar beet (Beta vulgaris), and wheatgrass (Thinopyrum intermedium), we demonstrate a potential efficient separation of three crossover (CO) processes on chromosomes: CO repulsion (positive COI), CO clustering (negative COI), and independent CO events (no COI).Keywords: Genetic recombination, Crossover interference, Coefficient of coincidence, Extension of the gamma-sprinkled model, Negative crossover interference Introduction Crossover interference (COI) refers to the non-random distribution of CO events (chiasmata) along chromosomes during meiosis. COI is reflected in deviations of the observed distribution from the expected pattern under the hypothesis of COs independence. In other words, COI leads to a lower or higher rate of double COs in adjacent regions than expected assuming no interference between COs. The phenomenon was discovered in the early experiments on recombination in Drosophila melanogaster [1–3]. The degree of COI can be quantified by the coefficient of coincidence (CoC), the ratio of the observed to the expected number of double COs for a pair of chromosomal intervals: CoC < 1, CoC >1, and CoC = 1, corresponding to positive, negative, and no COI, respectively. The nature of COI was explained in two ways: (i) as a result of interaction, either repulsion or clustering, of COs (referred to as chiasmata interference) [4], and (ii) as dependent involvement of chromatids in neighboring CO exchanges (referred to as chromatid interference) [2]. Studies of many species [5–11] have found no significant impact of chromatid interference on COI. Therefore, most COI models assume no chromatid interference [8, 12–14]. However, analysis in recent years of data from various organisms and mathematical modelling suggests that chromatid interference may exist, at least in some species [15, 16]. Since accounting for chromatid interference effects requires cytological data on chromatids involvement (which is currently available for very few species), and our datasets do not contain such information, this factor is not included into the current modeling framework. Future extensions of crossover interference models will likely require correction coefficients for interference parameters (such as CoC and parameters of different COI models) calculated using methodology [16], when appropriate cytological data become available. To describe CO interference as a single mechanism of CO formation operating at the chromosome or even the entire genome level, one of the most convenient formulations is the gamma model [12, 17–20]. It considers the dependence of the number of COs per gamete or the distances between COs along a chromosome on the interference “strength” parameter (v). The values v >1 correspond to positive COI (“repulsion” between COs), while v = 1 implies no interference. The gamma model was further modified in accordance with the discovery of the interference-independent metabolic pathway of meiotic recombination along with the common interference-dependent pathway [21]. The resulting “gamma-sprinkled” (GS) model is a mixture of two gamma distributions, with v >1 and v = 1 (corresponding to positive COI and no COI) with proportions 1-p and p, respectively, appeared to be more consistent with empirical data [22]. The case of CO clustering along the chromosome (i.e. negative COI) can be modelled by gamma-distribution with v < 1. Cases of possible negative COI (NCOI) in some genomic regions were reported in different organisms (Neurospora, yeast, Drosophila, Arabidopsis, maize, barley, wheat) [23–36]. Applying to such cases a standard GS model may result in biased estimates of both parameters, p and v, due to the presence of three rather than two components, with v >1, v = 1, and v < 1. Therefore, to account for possible NCOI, we extend the standard GS model by fitting a mixture of the corresponding three distributions. To date, no specific molecular mechanisms underlying NCOI have been identified. The phenomenon of NCOI is primarily regarded as a statistical model-based outcome, which may arise from a combined action of the three established meiotic crossover pathways [37, 38]: interference-dependent (ZMM-dependent/class I) and two interference-independent (class II/MUS81-dependent and class III/FANCD2-dependent [39]) pathways. Furthermore, novel coarsening models implicate HEI10 protein dynamics as a major regulator of crossover designation and interference patterns through diffusion-mediated accumulation mechanisms [40, 41]. Bearing in mind the existing solid experimental evidence on NCOI in different organisms (above cited [23–36]), in this study NCOI is treated as a component within the extended GS model framework, enabling the identification and quantitative description of regions with crossover clustering, despite the current absence of direct mechanistic explanations. Materials and methods Statistical analysis of COI landscape using CoC calculations on high-density linkage maps In general, for a pair of adjacent intervals on a genetic map flanked by three points (genetic markers, m1-m2-m3), the log-likelihood function for a testcross progeny can be presented as: where pij are the probabilities of the classes of gametes: p00 = (1-r1-r2 + CoC·r1·r2) for no COs on both intervals, p10 = (r1-CoC·r1·r2) and p01 = (r2-CoC·r1·r2) for CO on the 1 st and 2nd intervals, respectively, and p11 = (CoC·r1·r2) for double COs; while r1= and r2= are the recombination rates for the intervals m1-m2 and m2-m3 (calculated for each interval independently), and nij are the observed numbers of cases for the corresponding classes. With typical modern datasets consisting of a large number of markers and a relatively small sample size, using too short intervals implies insufficient statistical power to distinguish between the alternatives H1(CoC ≠ 1) versus H0(CoC = 1). Short intervals (< X cM) and long intervals (> Y cM) where X and Y are determined empirically according to the population size, genetic markers density on map etc. Namely, minimal size of the considered intervals is chosen as a function of the populations size allowing to observe singe crossover events. Likewise, we limit max length of the considered interval to avoid double crossovers that could be precepted as no recombinants. To further address this problem, we apply the “superposition” principle. Namely, increasing the interval size by using only a subset of markers is wasteful and may lead to the problem of classifying a genotype as “CO” or “non-CO” for longer intervals, because genotypes with double COs in such intervals will be classified as “non-COs”. A slight modification of the “CO genotype” definition for each new interval solves this problem. Specifically, we consider a genotype as recombinant in an extended interval if it was formed because of one or more CO events within the corresponding subintervals (considering all internal markers). We refer to this type of classification as the “crossover superposition principle” (or CSP). Accounting for heterogeneity in the COI landscape sliding windows on high-density linkage maps (illustrated with Norway Spruce real data) To estimate CoC, we maximize the likelihood function [42] using the simplex method [43, 44] for three hypothetical situations: (1) according to Haldane mapping function, i.e. with CoC = 1 [45]; (2) with COI according to the Kosambi [46] mapping function, i.e. with CoC = 2r; and (3) with CoC differing from both Kosambi and Haldane models and estimated directly from the analyzed (simulated or real) data. As a real-world example for illustration, we use the high-density linkage data for Norway spruce (Picea abies) published by Bernhardsson [47] (their data on segregating population, cluster #3 with n = 842 test-cross genotypes). For the analysis of interference, we constructed high-quality genetic maps using the MultiPoint-ultradense software [48]. To classify the genotypes into groups 00, 01, 10, and 11, we used the described in the previous section approach utilizing sliding windows along consequent pairs of adjacent intervals of varying sizes (l1 and l2) where l₁+l₂ represents the total genetic size of the current window. The obtained vectors of frequencies P = (p00, p01, p10, p11) = P(l1, l2) were employed for local maximum likelihood estimation of CoC for the sliding windows along the linkage map. We used the likelihood-ratio test [49] to compare the alternative COI models in each run (with an increasing sliding window size) followed by the Benjamini-Hochberg procedure [50] for false discovery rate correction. On the linkage group LG7 (according to the authors’ designations) we found a region with relaxation of positive COI followed by no-COI and even NCOI at some map points (Fig. 1). Fig. 1 demonstrates a comprehensive approach to COI landscape analysis along the genetic map, where model classification criteria are based not only on raw CoC estimates but also on statistical confidence intervals surrounding standardized interference models. This methodology employs LR testing to determine whether the observed CoC values differ significantly from theoretical models (Haldane, Kosambi) within specified confidence thresholds (α_χ2). The visualization reveals regions with stable model membership through color-coded zones: blue regions indicate “no-COI”, where the observed CoC does not significantly differ from the Haldane model (CoC ≈ 1), orange-red regions represent positive interference zones (Kosambi model or stronger, CoC < 1), and dark blue regions indicate negative interference (CoC >1). Statistically significant deviations are explicitly marked as SDPI (Significant Difference Positive Interference) and SDNI (Significant Difference Negative Interference) zones, allowing objective assessment of chromosome-wide interference landscapes. The sliding window approach with variable lengths (l1 + l2) addresses the analytical challenge where individual map positions may exhibit different COI classifications depending on the window size, particularly in transition zones between interference domains. This heat map presentation integrates results from multiple window sizes simultaneously, providing a robust framework for identifying local interference features independent of analytical window specifications and revealing the heterogeneous nature of recombination landscapes across chromosomal regions. A candidate place with significant NCOI was detected and represented by sliding windows of 14–15 cM length. These windows were formed by triplets of markers with different coordinates: left – at 79 cM, central – 82.1–83.5 cM, and right – 94.2–94.9 cM; with mean CoC = 2.39 ± 0.68 (cumulative p-value = 0.013, calculated by Fisher’s (1932) combined probability method with permutation correction to account for correlations between tests [51–53]. The permutation test was run with 1000 iterations, where haplotypes were randomly shuffled to generate an empirical distribution of the pooled statistics. The analysis of this region of the map with an increased sliding window size resulted in p-values that did not allow us to distinguish between the obtained CoC >1 and the Haldane model (i.e., no COI); yet, positive interference was not detected. This case may serve as an example of heterogeneity of local chromosomal regions for crossover interference. To estimate CoC, we maximize the likelihood function [42] using the simplex method [43, 44] for three hypothetical situations: (1) according to Haldane mapping function, i.e. with CoC = 1 [45]; (2) with COI according to the Kosambi [46] mapping function, i.e. with CoC = 2r; and (3) with CoC differing from both Kosambi and Haldane models and estimated directly from the analyzed (simulated or real) data. As a real-world example for illustration, we use the high-density linkage data for Norway spruce (Picea abies) published by Bernhardsson [47] (their data on segregating population, cluster #3 with n = 842 test-cross genotypes). For the analysis of interference, we constructed high-quality genetic maps using the MultiPoint-ultradense software [48]. To classify the genotypes into groups 00, 01, 10, and 11, we used the described in the previous section approach utilizing sliding windows along consequent pairs of adjacent intervals of varying sizes (l1 and l2) where l₁+l₂ represents the total genetic size of the current window. The obtained vectors of frequencies P = (p00, p01, p10, p11) = P(l1, l2) were employed for local maximum likelihood estimation of CoC for the sliding windows along the linkage map. We used the likelihood-ratio test [49] to compare the alternative COI models in each run (with an increasing sliding window size) followed by the Benjamini-Hochberg procedure [50] for false discovery rate correction. On the linkage group LG7 (according to the authors’ designations) we found a region with relaxation of positive COI followed by no-COI and even NCOI at some map points (Fig. 1). Fig. 1 demonstrates a comprehensive approach to COI landscape analysis along the genetic map, where model classification criteria are based not only on raw CoC estimates but also on statistical confidence intervals surrounding standardized interference models. This methodology employs LR testing to determine whether the observed CoC values differ significantly from theoretical models (Haldane, Kosambi) within specified confidence thresholds (α_χ2). The visualization reveals regions with stable model membership through color-coded zones: blue regions indicate “no-COI”, where the observed CoC does not significantly differ from the Haldane model (CoC ≈ 1), orange-red regions represent positive interference zones (Kosambi model or stronger, CoC < 1), and dark blue regions indicate negative interference (CoC >1). Statistically significant deviations are explicitly marked as SDPI (Significant Difference Positive Interference) and SDNI (Significant Difference Negative Interference) zones, allowing objective assessment of chromosome-wide interference landscapes. The sliding window approach with variable lengths (l1 + l2) addresses the analytical challenge where individual map positions may exhibit different COI classifications depending on the window size, particularly in transition zones between interference domains. This heat map presentation integrates results from multiple window sizes simultaneously, providing a robust framework for identifying local interference features independent of analytical window specifications and revealing the heterogeneous nature of recombination landscapes across chromosomal regions. A candidate place with significant NCOI was detected and represented by sliding windows of 14–15 cM length. These windows were formed by triplets of markers with different coordinates: left – at 79 cM, central – 82.1–83.5 cM, and right – 94.2–94.9 cM; with mean CoC = 2.39 ± 0.68 (cumulative p-value = 0.013, calculated by Fisher’s (1932) combined probability method with permutation correction to account for correlations between tests [51–53]. The permutation test was run with 1000 iterations, where haplotypes were randomly shuffled to generate an empirical distribution of the pooled statistics. The analysis of this region of the map with an increased sliding window size resulted in p-values that did not allow us to distinguish between the obtained CoC >1 and the Haldane model (i.e., no COI); yet, positive interference was not detected. This case may serve as an example of heterogeneity of local chromosomal regions for crossover interference. Similar landscape patterns have been found by many authors in pericentromeric regions [24, 27, 29, 54–57]. Similar landscape patterns have been found by many authors in pericentromeric regions [24, 27, 29, 54–57]. COI landscape along the genetic map of Norway spruce LG7 (based on data from Bernhardsson [47]). The analysis is performed using variable-size sliding windows (l1 + l2) on the Y axis; X axis - genetic map position in centimorgans). The colored regions correspond to neighborhoods of coefficient of coincidence (CoC) values in accordance with their confidence intervals: If the observed CoC does not significantly differ from the Haldane model, it is classified as “no-COI” (blue). When CoC appears in the robust zone around Kosambi or stronger positive interference it is colored in orange or red, respectively. If CoC is significantly higher than 1 and differs from Haldane model, negative interference is recorded (dark blue). The plot also shows schematically how regions are classified and the statistical criteria for assignment (likelihood ratio), comparison to threshold values with LR-score. Zones of statistical significance (SDPI and SDNI — significant difference for positive and negative COI, respectively) are highlighted. The diagram also illustrates the problem of projecting COI landscape analysis onto a genetic map using sliding windows of varying lengths, where a single map point can correspond to different COI models in border transition zones depending on the sliding window size used for the analysis. This is why; to visualize the heterogeneous recombination landscape in the figure, we use a heat map for all runs of variable-length windows analysis simultaneously In order to characterize the chromosomal landscape of COI, we also performed a chromosome-wide analysis using simulations with CODA software [56], allowing us to estimate the parameters of the standard gamma-sprinkled model (Fig. 2). The solution obtained based on simulations points to the presence of a non-interfering component in the distribution of recombination events in the considered data, with a proportion p = 0.29. Likewise, the data showed an excess of gametes with CO number >3 (Fig. 2, right) and an excess of short inter-crossover (CO-CO) distances (Fig. 2, left). The observed deviation of the CODA simulations with the obtained estimates of v and p parameters, i.e., the excess of both multiple exchanges and short inter-CO intervals, remarkably coincides with our finding of possible negative COI for the same considered linkage group (see Fig. 1). However, a more objective conclusion about how to interpret these deviations from the standard GS model requires an extension of the GS model allowing for the possibility of negative interference. CODA simulations using the GS model for LG 7 Norway spruce data Bernhardsson [47] Left - CO-to-CO distance distribution for gametes with CO number >2; right - CO number per gamete distribution. The obtained estimates of GS parameters: v = 7.90, p = 0.29; green columns – real data; red line – simulation Extension of the GS model to account for negative COI COI model based on gamma distribution Let crossover events occur on the chromosome with a distance from the previous CO event, depending only on the position of this previous event. Let the distances D on the genetic map between crossover events follow a gamma distribution with parameters v (shape) and θ (scale) [18]: This model can be viewed as a generalization of the classical Haldane model, where CO events are independent so that the distances between successive events follow an exponential distribution: From Γ(1, 1/𝜆)~𝐸𝑥𝑝(𝜆) follows the Haldane model [45]. If 2v is an integer, then the gamma distribution coincides with χ2 distribution with 2v degrees of freedom (v = df/2, 𝜃 = 2): With df = 4, v = 2, and 𝜃=100/2, the gamma distribution coincides with the χ2 distribution with df = 2. Such distribution of distances between COs is close to the Kosambi model [46]. In general, the average and variance of the distances between consecutive COs fitting the gamma model can be presented as: and where c is a scaling coefficient. Genetic distance (in Morgans) is the expected number of CO events between two points on the genetic map; in terms of ‘cM distances’: E(D) = 100 = vθ. Hence, θ = 100/v and to specify the model we need only parameter v. The case with v = 1 corresponds to the situation of no COI, v > 1 implies positive COI, and 0 < v < 1 - negative COI. Based on the symmetry, the doubled distance from the chromosome “begin” to the first CO. events should also follow the model (2). Hence, the distribution of this distance can be presented as: Simulation of distances between two CO events based on gamma distribution with parameter v Define 𝐺(𝑥) and 𝐹𝑣(𝑥) as distribution functions of the genetic distances from the “beginning” of the chromosome to the first CO event and from the position of the observed CO to the next CO under the gamma distribution model with the corresponding shape parameter, respectively: Let u be a random variable with uniform distribution on the interval (0, 1): U(0,1). We simulate Dv as 𝐷𝑣 ≔ 𝐹𝑣−1(𝑢). For calculations of Fv(x), we use the following approximation: Note that . Therefore, this approximation is problematic for v<1 and small xi. In such cases, we can use the following asymptotic: We simulated CO events on a chromosome with genetic length L (scored in cM) based on Eqs. 8–10 so that the CO map positions are values from [0, L]. Let u(i) ~ U[0,1], 𝐷(0) ≔ 𝐹1 − 1(𝑢(0)), 𝑥(0) ≔ 0.5 𝐷(0). Then. We stop the simulation when 𝑥(𝑗) + 𝐷(𝑗+1) > 𝐿. The resulting ordered set of values 𝑥(𝑖) is denoted by vector X. Simulation of CO events based on a mixture of three processes (with parameters v0, v1, and v2) Let three types of CO events obey the models with shape parameters v0, v1, and v2, while events of one type do not depend on events of other types. Let p1 and p2 be the proportions of events of types 1 and 2 on the chromosome, 0 ≤𝑝1, 𝑝2 ≤ 1, and 𝑝1 + 𝑝2 ≤ 1. If 0 < p1, p2, p1+p2 < 1, then to maintain on average 0.01 CO rate per 1 cM, the following condition should hold: 𝑝1𝑣1𝜃1 + 𝑝2𝑣2𝜃2 + (1 − 𝑝1 − 𝑝2) 𝑣0𝜃0 = 100, so that Our simulations are based on iterations with . Simulations continue while . We unite the three sets of the obtained values and sort them in increasing order. The resulting vector is denoted by X. Simulation of CO events with changing parameter v along the chromosome Let parameter v vary along the genetic map: v = v(x). Define . The distribution density of the distance to the next CO event (after the CO event at x0) is. To simplify the calculations, we can use the following equation: By integration of this equation, we obtain the following formula for : Analogously to the case with constant v, we simulate positions of CO events on the map [0, L] as: , ; and , , where u(j)~U[0,1], . Simulation study demonstrating the application of the extended GS model As an illustration of the application of the described approach, we generated a testcross dataset with characteristics similar to the Norway spruce real data mentioned above (Sect. 3). We simulated several populations (with n = 200, 500, 1000, and 5000) encompassing chromosome zones with varying v, including strong positive COI (v = 5), Kosambi COI model (v = 2), Haldane COI model (v = 1; no COI), negative COI (v = 0.5), and strong negative COI (v = 0.2). We analyzed the distribution of CoC using two approaches: the “crossover superposition” principle CSP (described in Sect. 3) and standard classification (based on flank markers). For population sizes N = 200 and 500, we employed window sizes l1 + l2 = 20 cM and 30 cM, while 10 cM and 20 cM windows were employed for N = 1000 and 5000, bearing in mind the increase in resolution with a higher sample size (Fig. 3). We found that for cases of strong NCOI, CoC estimates based on the CSP classification are more accurate than those obtained from the standard classification. This underscores the efficacy of the CSP approach in improving the accuracy of CoC estimates, especially in the context of strong NCOI. This approach will be most effective due to the ever-increasing trends in improving the quality of the genetic mapping data (including higher marker density and population size). Our simulations confirmed that for the LG7 Norway spruce data considered in Sect. 3 (population size n = 842, map length ~ 150 cM), using a sliding window size of ~ 15 cM, we could detect a robust zone of negative COI with a possible transition to no COI. A notable drawback of employing CoC calculations to assess local interference heterogeneity lies in their sensitivity to the interval length, population size, and the local marker density in the target region of the genetic map. To characterize the effect of NCOI along the genome, between different genomes, and sexes, more effectively would be to employ a model allowing simultaneous NCOI component alongside the “classical” positive COI and no COI. Simulation of COI landscape with variable v. Simulated v is represented by a step blackline, variation in CoC estimates based on flanked-markers and CSP genotype classification are shown by solid and dotted lines, respectively, for each size (w = 10, 20, 30 cM) of the sliding window Statistical analysis of COI using the extended GC model Basic statistics for a single vector and a set of vectors of CO positions Let X be a vector of consequent CO positions on a chromosome for an individual of a segregating population. The dimension of this vector(dim(X) ≥ 0) corresponds to the number of CO events for the individual. Let further denote the set of lengths between consecutive CO events by. 𝑣𝑙𝐼𝑛𝑡𝑃𝑎𝑟𝑡(𝑋) = {𝑥𝑖+1 − 𝑥𝑖|𝑖 = 1,2, …, dim(𝑋) − 1}; if dim(X) ≤ 1 then the set 𝑣𝑙𝐼𝑛𝑡𝑃𝑎𝑟𝑡(𝑋) is empty. Let 𝑿 be the set of vectors X for the entire population. Then we denote by 𝑁𝑚(𝑿) the number of vectors X from set 𝑿 such that dim(X) = m, and is the proportion of vectors X with dimension m and is a vector of the observed proportions of individuals with different numbers of COs. We unite all cases with 𝑚 ≥ 5 because such cases are extremely rare in most genetic mapping studies [56]. We denote the observed cumulative distribution function of distances between consequent COs within such subset of 𝑿 as Distribution of the considered statistics For a given chromosome length L and vector of parameters , we generate a large population 𝑿𝑉of N independently simulated gametes (e.g.,N = 5000).). Each gamete Xi is represented by a list of CO map positions generated according to the method described in paragraph 4.3. Based on set 𝑿𝑉 we estimate distributions 𝒑(𝑉)= 𝒑(𝑿𝑉) and 𝐹(𝑉)(𝑥) = 𝐹(𝑥, 𝑿𝑉). Testing the model adequacy To compare alternative models for the test data, we employ the statistics p and F described above and calculate statistical support (p-values) using standard criteria: the χ2-test and Kolmogorov-Smirnov test. To calculate these statistics, first, we need to estimate map positions of all CO events observed on N individuals. Then we estimate p, based on the observed proportions of individuals with different numbers of COs, and F, based on the observed distribution of CO-CO distances scored on the subset of individuals with exactly two COs. These statistics are used to compare two alternative hypotheses for the considered dataset - H0: {the data are consistent with the model with the defined parameters}, and H1: {the data can be explained by some alternative model}. For simplicity, we only discriminate between models described in Sect. 4 with different sets of parameters V = (v0, v−, v+, p0, p−, p+), i.e., H0: 𝑉 = 𝑉(0) versus H1: 𝑉 ≠ 𝑉(0). We specifically focus on comparing the null hypothesis of the absence of COI with different alternative hypotheses. These alternatives include scenarios with only positive COI, only negative COI, a combination of positive COI and no-COI, a combination of negative COI and no-COI, and a mixture involving all three distribution components. For example, this way we can distinguish between the GS model (𝑝−=0, i.e., no negative COI) and our extended GS model, which allows for some negative COI. To discriminate between these hypotheses, we use the likelihood ratio test. Instead of the quite complicated likelihood function for the set 𝑿, we used likelihood functions for two simple statistics, p and F(x), of this set. The likelihood for the statistic p can be calculated based on the distribution function for the statistics in the χ2-test to compare two discrete distributions. The likelihood for the statistic F(x) can be calculated based on the distribution function for the test statistic in the Kolmogorov-Smirnov test for the comparison of two continuous distributions [58]. χ2 test based on the observed distribution of CO numbers: this test assesses the differences between the observed and expected proportions of CO number classes per chromosome. The test statistic is computed using the distributions 𝒑(obs) and 𝒑(𝑉) expected under model V. Under H0, the test statistic X2 should follow a χ2 distribution with 5 degrees of freedom (since we employed 6 categories for the observed CO number). The p-value is calculated as the probability of observing a value of X2 that is as extreme or more extreme than the observed value: 𝑝 = Pr( ≥ 𝑋2). For the defined set of the model parameters V, the distribution 𝒑(𝑉) can be estimated by intensive machine simulations. The likelihood function based on statistic p can be defined by. The χ2 test has limitations when applied to data with a small number of categories, such as in cases of short chromosomes where the average number of crossovers is low. In such cases, the assumptions underlying the χ2 test may not hold, leading to unreliable results, so it is important to use an additional test. Kolmogorov–Smirnov test: this test compares the cumulative distribution functions (CDFs) (obs)(𝑥) = 𝐹 (𝑥, 𝑿) and 𝐹(𝑉)(𝑥) of the observed and expected (under the model with the vector of parameters V) distances between chromosomal positions of CO events in individuals. To assess the adequacy of our model, we employed a method that handles separately groups of individuals with any fixed number of crossovers per considered chromosome (two, three, four, five, and more). For chromosomes with n crossovers, we analyzed the (n−1)-dimensional vector of map distances between consecutive crossovers. First, we used a multivariate generalization of the classical Kolmogorov–Smirnov test proposed by Fasano and Franceschini [59]. this method extends the univariate test by constructing hyper-quadrants around each data point and comparing the fraction of points falling within each region across the two samples. this approach is particularly well-suited for our analysis as it handles multivariate data without assuming any specific parametric distribution - an important consideration given that crossover interference can induce complex, non-Gaussian dependencies between inter-crossover intervals. In our implementation, we used the R package fasano.franceschini.test [60]. To further validate our findings and account for different statistical sensitivities, we also employed the test of Biau and Györfi [61], available in the DataSimilarity R package [62]. This test assesses the homogeneity of two multivariate distributions based on the L1-distance between histogram-based density estimates. A key strength of the Biau–Györfi test is its robustness to the “curse of dimensionality” through its reliance on adaptive partitioning, allowing it to detect differences in distributional shape even in moderate-dimensional data. We evaluate its performance via permutation testing, which generates an empirical null distribution to calculate accurate p-values. Since the p-values for the considered groups with different numbers of crossovers (2, 3, etc.) are dependent, to obtain the cumulative p-value, we use the previously mentioned Fisher’s [49] combined probability method with permutation correction to account for correlations between tests [51, 52]. This comprehensive, nonparametric, and multidimensional testing framework strengthens our ability to rigorously evaluate the fit of the extended GS model across diverse recombination scenarios. The test statistic represents the maximum discrepancy between F(obs)(x) and F(V)(x). Under certain assumptions, this statistic follows the Kolmogorov distribution. Thus, the p-value is computed as: , where FK(x) is the cumulative distribution function for the Kolmogorov distribution. The likelihood function based on the statistic F(x) can be defined as: Limitations for applying the extended GC model to real data The motivation of our approach is to estimate the contribution of the three putative mechanisms of CO localization (COI-free, positive COI, and negative COI) to the observed patterns of multilocus recombination in real data. We hypothesize that the three mechanisms may operate simultaneously in the same chromosome or even chromosomal region. Thus, we model such situations by a mixture of three processes (Sect. 4), with parameters v0 = 1 (no COI), 0 < v−<1 (NCOI), and v+>1 (positive COI). For different vectors (v0, v−, v+, p0, p−, p+) we calculate the significance level for CO-number and CO-CO distances based on statistical tests (Sect. 5.3). We consider only models with p-values not higher than 0.05 threshold for both statistical tests to be a relevant description of the analyzed data. Such an approach enables estimating model parameters assuming homogeneity of the entire process along the chromosome or large enough chromosomal segments. We simulated a testcross progeny with 2000 genotypes and 200 markers on a chromosome of 200 cM length, with the following parameter values: v0 = 1, v−=0.5, v+=2, p−=0.2, p+=0.4, p0 = 1-p−-p+=0.4. This combination of parameters was chosen assuming a situation where moderate positive interference (v+=2) is mixed with no-COI (v0 = 1) in equal proportions (40%), and there are also 20% events affected by intense negative COI (v−=0.5). Using the described procedure (Sect. 5.1–5.3), the vector of maximum likelihood parameter estimates (maximizing the min {𝐿𝒑 (𝑿, 𝑉), 𝐿𝐹 (𝑿, 𝑉)} functional) was: =0.5, =1.9, =0.25, =0.45. We tested hypothesis H1 that parameter values are equal to the estimated ones against the alternative H0 = {p−=0, p+=1} (i.e., assuming unimodal distribution with the presence of only positive COI with =1.10). The distribution of CO numbers in our simulated data significantly (p-value < 0.001) differed from the one expected under H0. Therefore, the proposed analysis can unravel the mixed nature of interference patterns. Following the previous section, we compared our “extended” GS model (v0, v−, v+, p0, p−, p+) with the standard GS model (v+, v0, p+, p0). Our objective was to gain insights into the influence of complex relationships between chromosome length, population size, and relative proportions of the three mechanisms on the outcome of COI models comparison. To ensure the rigor and robustness of the analysis, we conducted a comprehensive investigation using Monte-Carlo simulations. The simulations were performed for each scenario, wherein we systematically varied the population size. Furthermore, to explore the effect of different genetic map lengths, we assessed various scenarios, spanning from 20 to 400 centimorgans (сМ) (Table 1). The main objective was to evaluate the relationship between population size and the power of the test (1-β). A similar approach was applied to sets of chromosome lengths, with population size serving as the independent variable. In each simulation, three critical p-value thresholds were considered (0.1, 0.05, and 0.01). Table 1Comparison of the standard GS model and the extended GS model (v0, v-,v+, p0, p-,p+) Chromosomelength (cM)Samplesizealpha level(α)Chromosomelength (cM)Samplesizealpha level (α)0.10.050.010.10.050.01The test power (1-β)The test power (1-β)20500.090.060.03120500.290.180.061000.140.10.051000.470.320.142000.260.20.132000.750.610.323000.350.280.193000.880.770.55000.520.430.325000.970.940.797500.690.620.477500.990.980.9410000.80.710.58100010.990.9820000.980.960.88200011150500.170.10.03150500.260.160.041000.310.210.091000.480.320.122000.550.430.212000.750.590.33000.720.590.363000.870.770.55000.910.820.625000.970.930.777500.960.940.847500.990.980.9310000.990.980.931000110.972000111200011175500.240.150.04200500.250.140.041000.410.30.121000.450.290.092000.650.510.272000.70.520.243000.840.720.483000.810.70.395000.960.920.765000.930.860.657500.990.980.927500.960.950.841000110.9810000.980.970.9120001112000110.99100500.250.160.05400500.230.140.031000.440.310.121000.490.290.12000.710.550.282000.750.610.293000.860.750.463000.880.780.515000.960.910.755000.940.910.777500.990.980.927500.970.960.911000110.9810000.980.980.9520001112000110.99 Comparison of the standard GS model and the extended GS model (v0, v-,v+, p0, p-,p+) Our numerical tests showed that for achieving an acceptable level of test power in comparisons between different COI models (even in situations of high marker density) and distinguish between the standard GS model (v+, v0, p+, p0) and the extended GS model (v0, v−, v+, p0, p−, p+), samples of more than 150–200 individuals would be more relevant. Notably, the power of the test to distinguish between the COI models was influenced more by the chromosome length than the population size. This suggests that researchers should collect larger samples when conducting COI analysis for species with short genetic maps. Results: analysis on real data with extended gamma sprinkling model To assess the performance of our modified GS model, we selected a few suitable publicly available datasets on plants: wheatgrass [63], conifer larch [64], and sugar beet [65]. Using a sliding window and CoC calculation method, we first looked for areas of negative COI in the COI landscape along the chromosome (Table 2). Table 2Zones of negative COI were detected with a sliding windowOrganismLinkage mapSample sizeMarker groupsPositions on genetic map (cM)Left flankCenterRight flankWheatgrassLG7 female264[15 - 21] [21, 25][26 - 29]56.1 - 67.567.5 - 74.481.6 - 93.1Average CoC = 3.21 ± 0.89, zone p -value =2.76×10−25LG9 female264[0, 10][2, 12, 14][15, 16]0.0 - 57.87.2 - 72.819.7 - 82.9Average CoC = 2.12 ± 0.76, zone p -value =0.033LG18 male342[1 - 8][3 - 9][8 - 11]28.5 - 78.856.3 - 84.768.6 - 98.3Average CoC = 3.32 ± 1.47, zone p -value =0.0012LarchFemale 10144[1 - 15][8 - 27][14 - 42] 0.0 - 49.917.0 - 50.625.4 - 73.3Average CoC = 5.16 ± 1.91, zone p -value =0.1×10−25Sugar beatC1 female151[1, 2][2, 3, 4][3 - 10] 52.4 - 109.374.5 - 123.5 92.8 - 142.4Average CoC = 1.63 ± 0.28, zone p -value =2.56×10−18 Zones of negative COI were detected with a sliding window [15 - 21] [21, 25] [26 - 29] [0, 10] [2, 12, 14] [15, 16] [1 - 8] [3 - 9] [8 - 11] [1 - 15] [8 - 27] [14 - 42] [1, 2] [2, 3, 4] [3 - 10] We selected chromosomes on which we identified “robust” zones of NCOI with high CoC > 1, represented by groups of flanking left (L) and right (R) markers and central (C) markers of the sliding window. The C group of markers showed a pattern of heterogeneity in the interference landscape on the chromosome in the form of a NCOI “valley”, similar to the model described earlier in Sect. 4.4: weakening of positive COI along the genetic map to the absence of COI, followed by the appearance and strengthening of NCOI (see Fig. 3). These data fit the extended GS model (v0, v−, v+, p0, p−, p+). For the subsequent analysis, we found it appropriate to normalize the model parameters to account for the simultaneous presence of two COI components with opposite deviations from the no-COI pattern. We associate the extended GS model (v0, v−, v+, p0, p−, p+) with а singe-shape parameter v (either < 1, =1, or > 1) of gamma model. This normalization allows for the comparison of various mixture models on a common scale. Such simplification facilitates the evaluation of the relative contribution of clustering and repulsion components. We propose the following normalization for the shape parameter: This formula accounts for different scenarios: (i) = = 0, =1; (ii) = 0, = 1, = ; and (iii) = 0, = 1, . Correspondingly, the normalized parameters are defined as: The results of this normalization and application of the extended GS model to the analyzed chromosomes (Table 2) are presented in Table 3. Table 3GS model (v0, v-, v+,p0, p-, p+) with normalization of the shape parameter vSpeciesLinkage groupExtended GS model (v 0, v -, v +, p 0, p -, p +)Normalized valuesv +v -p +p -p0%WheatgrassLG7 female100.9406004.600.94LG9 female1.20.72070101.040.79LG18 male50.73040301.620.86LarchLG10 female1.10.63020501.020.90SugarbeatC1 female1.20.83020501.030.96 GS model (v0, v-, v+,p0, p-, p+) with normalization of the shape parameter v In our formula for the model normalization, the denominator 𝑣𝑚𝑎𝑥+ − 1 allows to remove the imbalance of 𝑣− and 𝑣+ ranges. The best value for 𝑣𝑚𝑎𝑥 can be estimated from simulated data for each specific experimental situation. The presented algorithm is implemented in software tools (https://github.com/stingerfun/Extended-Gamma-Sprinkled-Model). Discussion Variation of crossover interference along a genetic map can be investigated at local and chromosome-wide levels. Both these approaches have limitations. In practical situations, COI can vary along the chromosome, resulting in low compliance of the chromosome-wide model to real data. For a qualitative assessment of the local COI, the standard analysis using CoC strongly depends on the population size and the length of the map intervals under study. To address the common problem of insufficient sample size for COI detection, we propose, as an option, to employ larger size intervals. With this approach, we consider an individual as ‘recombinant’ in a larger-length interval if one or more CO (either odd or even) events have been registered in the sub-intervals of the larger interval. We demonstrated that CoC estimates for such classification of genotypes have a certain advantage over the standard classification. This makes it possible to better assess the impact of negative COI on the local scale of the genetic map. To check for the robustness of the derived conclusions, we perform the analysis on a range of sizes of the consequent intervals along a chromosome. For the chromosome-wide analysis of COI, the currently widely employed GS model (v+, v0, p+, p0) does not allow testing for the presence of negative COI, which may exist as demonstrated by some empirical evidence [23–36, 66–71]. Due to the absence of such a possibility in the GS model (v+, v0, p+, p0), negative COI cases (when present) are mixed with no-COI cases in the output results. In this paper, we proposed an extension of the GS-model aimed at considering cases of possible (local) negative COI. For this purpose, we analyzed the real data by fitting the whole distribution of crossover events as a mixture of three rather than two COI processes (positive, negative, and no-COI). In this way, we estimate distributions of the CO number per genotype and inter-CO distances expected under this model. Comparisons of the distributions observed in the analysis of real data with the massive simulations enable estimating the parameters of the model and testing the corresponding hypotheses. This approach should assist in more objectively addressing the question of the presence of negative COI in each suitable experimental data, rather than ignoring it outright. Usually, COI is considered a factor preventing the formation of excessive COs [38]. Recent studies in Arabidopsis and some crops have shown that loss-of-function mutations in anti-recombination genes involved in DNA repair of excessive double-strand breaks during meiosis can result in a very high increase in CO rates [68–71]. What is the role of deregulation of COI in this increase remains an open question. It is important to mention that in some earlier reports, including drosophila and wheat, negative COI was remarkably associated with pericentromeric regions [24, 27, 29, 55, 56], where normally recombination is strongly suppressed. Acknowledgements The reported study was supported by the Israel Science Foundation (Grant/Award Number: #1844/17); the US-Israel Binational Agricultural Research and Development Fund (Grant/Award Number: IS-5116-18); the Graduate Studies Authority of the University of Haifa; and the Israeli Ministry of Aliyah and Integration. Abbreviations CentiMorgan Crossover Crossover interference Coefficient of coincidence Software for crossover distribution analysis Crossover superposition principle Degrees of freedom Gamma-sprinkled Linkage group Negative crossover interference Authors’ contributions SS and ZF contributed equally to this work. SS, ZF, EP and ABK designed the study. SS and ZF developed the methodology and performed the analyses. SS implemented the software. SS, ZF and ABK wrote the manuscript. All authors read and approved the final manuscript. Funding The reported study was supported by the Israel Science Foundation (Grant/Award Number: #1844/17); the US-Israel Binational Agricultural Research and Development Fund (Grant/Award Number: IS-5116-18); the Graduate Studies Authority of the University of Haifa; and the Israeli Ministry of Aliyah and Integration. Data availability The software tools developed in this study are available at: https://github.com/stingerfun/Extended-Gamma-Sprinkled-Model. The datasets analyzed during the current study are available from the corresponding authors on reasonable request. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Shaul Sapielkin and Zeev Frenkel contributed equally to this work. Contributor Information Shaul Sapielkin, Email: shaul@evo.haifa.ac.il. Abraham B. Korol, Email: korol@evo.haifa.ac.il References Associated Data Data Availability Statement The software tools developed in this study are available at: https://github.com/stingerfun/Extended-Gamma-Sprinkled-Model. The datasets analyzed during the current study are available from the corresponding authors on reasonable request.