The core challenge any model of genetic variation must confront is linkage disequilibrium, the nonrandom association of variants across the genome. Neighboring SNPs, or single nucleotide polymorphisms, tend to be inherited together, but so do many distant ones, and these long-range correlations carry much of the structure that makes a population recognizable in the data. Classical simulators built on the coalescent model capture these patterns by tracing observed variation back to latent genealogies shaped by demographic history, mutation, and recombination. That approach is expressive but computationally punishing, which is why practical tools such as msprime rely on Markovian approximations. A second tradition, descending from the product-of-approximate-conditionals framework of Li and Stephens, dispenses with explicit genealogies and fits the data distribution directly, naturally yielding hidden Markov models. HMMs have been enormously successful in haplotype phasing, genotype imputation, and ancestry inference, but their chain structure forces information between distant SNPs to pass through every intermediate hidden variable, weakening long-range dependencies at each step.

Deep learning promised a way out, and a wave of generative adversarial networks, variational autoencoders, restricted Boltzmann machines, and most recently diffusion models has been applied to genetic data. These models can produce artificial genomes that look plausible in principal component analyses and allele frequency summaries. Yet the UCLA-led team argues they carry structural liabilities for genomics. Generative adversarial networks do not define a probability distribution over the data at all, making likelihood-based inference impossible. Restricted Boltzmann machines define one, but evaluating it requires computing a partition function over an exponentially large space of configurations. Variational autoencoders expose only a lower bound on the marginal likelihood. Diffusion models can in principle compute exact likelihoods but scale poorly to dense SNP data and have not been evaluated on imputation. Crucially, none of these approaches supports efficient estimation of conditional probabilities, the mathematical operation at the heart of genotype imputation, and none offers an objective way to judge whether training has converged.

GPC’s answer is architectural. The model builds on hidden Chow-Liu trees, a class of latent variable models in which every observed SNP is paired with a hidden variable, and the hidden variables are wired together not in a chain but in a tree learned from the data itself. The Chow-Liu algorithm, dating to 1968, computes pairwise mutual information between all variables and connects them in the maximum-weight spanning tree, placing strongly correlated SNPs adjacent to one another regardless of their positions along the chromosome. Where a hidden Markov model would route the dependence between two distant but correlated SNPs through every latent variable in between, losing correlation at each hop, the learned tree carries it over a single edge. In the tree learned from a 10,000-SNP region of chromosome 15 in the 1000 Genomes dataset, the researchers found 4,599 leaf nodes, 397 nodes with degree greater than two, one hub connected to 161 other SNPs, and 18.3 percent of edges spanning more than 1,000 positions apart, a topology fundamentally unlike the unbranched chain of an HMM.

Expressiveness alone would be useless if the model were computationally intractable, and this is where probabilistic circuits enter. A probabilistic circuit represents a joint distribution as a directed acyclic graph of input, sum, and product nodes, with product nodes encoding factorizations and sum nodes encoding weighted mixtures. When the graph satisfies two structural properties, smoothness and decomposability, any marginal probability can be computed exactly in time linear in the circuit’s size. Conditional probabilities, the quantities needed for imputation, follow as simple ratios of two marginal queries, each evaluated in a single feedforward pass. Sampling artificial genomes proceeds by ancestral traversal of the circuit, also linear in size. By compiling hidden Chow-Liu trees into this circuit representation and training them with expectation-maximization on GPUs using the PyJuice package, the team trained models with more than 88 million parameters, with each EM epoch taking under two seconds on the 1000 Genomes data and full training completing in roughly two to six hours on a single NVIDIA RTX A5000 graphics card.

The practical payoff shows up most clearly in genotype imputation, the task of inferring variants a person was not directly genotyped, which underlies much of genome-wide association science. The researchers evaluated GPC across three settings using the 1000 Genomes Project, the UK Biobank, and a high-coverage 1000 Genomes release. In the general setting, where models trained on diverse ancestries impute into similarly diverse test genomes, GPC used for direct conditional imputation achieved a 27.5 percent average improvement in the squared correlation between imputed and true genotypes over the next best method, a restricted Boltzmann machine, and a striking 174 percent improvement for low-frequency variants with minor allele frequency below one percent. Direct imputation through the model itself also beat GPC’s own artificial genomes used as reference panels for the standard tool Impute5, by about 15.5 percent overall, apparently because it targets the imputation objective directly rather than injecting the noise of an intermediate simulation step.

The gains were largest precisely where existing infrastructure is weakest: populations underrepresented in public reference panels. Because large public datasets are overwhelmingly of European ancestry, imputation into African, admixed, or other non-European populations suffers from distributional mismatch. In population-specific experiments, GPC’s direct imputation improved on the next best deep generative method by 33 percent on average, and by 279 percent for low-frequency variants. Remarkably, in the 1000 Genomes experiments GPC trained on ancestry-matched private data outperformed Impute5 running on real European reference genomes, achieving a 12.3 percent average improvement in squared correlation and a 42.1 percent improvement for rare variants. In a realistic array-based scenario, imputing 12,551 SNPs missing from the HumanOmni5Exome genotyping array, GPC outperformed the European-panel Impute5 benchmark by 96.5 percent on average and by more than twelvefold for low-frequency variants. The entire held-out array imputation took about one second in a single conditional query.

Privacy, the original motivation for artificial genomes, also fared better under GPC. Using the nearest-neighbor adversarial accuracy metric, which scores values near 0.5 as the ideal balance between utility and privacy, GPC came closest to the ideal across nearly all splits of both datasets. The failure modes of the competitors were instructive. Restricted Boltzmann machines produced synthetic haplotypes that each sat closer to some individual real genome than to any other synthetic one, meaning samples clustered around training individuals, a memorization pattern that could allow synthetic genomes to be traced back to specific people. The Wasserstein generative adversarial network showed the opposite pathology, with artificial genomes occupying a region largely disjoint from the real data, sacrificing utility. The hidden Markov model failed in the most extreme form, producing genomes perfectly separable from real ones in both directions, achieving maximal privacy by failing to model the data at all. GPC’s artificial genomes were not perfectly indistinguishable, and the authors caution that improved performance on this metric does not guarantee protection against membership inference or attribute inference attacks.

Limitations remain, and the authors are candid about them. GPC currently operates on single genomic regions of roughly 10,000 to 15,000 SNPs, bounded by the memory cost of constructing the Chow-Liu tree, which grows with the square of the SNP count; reaching genome-wide scale will require hierarchical or distributed approaches. The present implementation handles haploid rather than diploid genomes, and like all generative models GPC inherits biases from its training data, leaving fairness across diverse cohorts an open problem. Still, the study marks a notable convergence of two research traditions that have mostly run in parallel: the tractable probabilistic models beloved of population geneticists and the expressive deep architectures driving modern machine learning. By learning the dependence structure of the genome directly from mutual information and then compiling it into a form where exact inference is cheap, GPC suggests that the trade-off between realism, computational tractability, and privacy in synthetic genomic data may be less inevitable than it once appeared. Code and experiments are publicly available as the field moves toward equitable genomic tools for populations long left out of reference panels.

Subject of Research: A tractable deep generative model, Genetic Probabilistic Circuits, for modeling human genetic variation data

Article Title: GPC: An expressive and tractable deep generative model for genetic variation data

Article References: Anand, P., Liu, A., Dang, M., Fu, B., Wei, X., Van den Broeck, G., & Sankararaman, S. (2026). GPC: An expressive and tractable deep generative model for genetic variation data. PLOS Genetics, 22(10), e1012321. https://doi.org/10.1371/journal.pgen.1012321

Image Credits: AI Generated

DOI: 10.1371/journal.pgen.1012321

Keywords: genetic probabilistic circuits, artificial genomes, genotype imputation, linkage disequilibrium, probabilistic circuits, hidden Chow-Liu trees, population genetics, deep generative models, genomic privacy, 1000 Genomes Project, UK Biobank, underrepresented populations

The concept of covariance measures how two random variables vary together, and it is the statistical glue that connects relatives, traits, and environments in classical theory. Within a single population, the covariance between relatives arises because they share alleles that are identical by descent. But when an individual comes from one biparental cross, say an indica rice cross A/B, and another from a completely different cross, such as a japonica cross C/D, there is no established framework for asking how much their genotypic values co-vary. Rebollo and Bernardo propose an elegant answer: shared marker alleles between individuals belonging to two different populations can serve as the basis for defining and estimating a between-population covariance. In other words, even when two populations share no recent ancestry, the molecular markers they have in common provide a measurable bridge for quantifying joint genetic variation.

The analytical core of the method extends earlier formulas developed by Osthushenrich and colleagues in 2017, which estimated the genetic variance within a recombinant inbred line population from marker effects. The authors define the genotypic value of an individual as a random variable whose variance can be decomposed across chromosomes and loci. Under a purely additive genetic model, the heterozygote effect is zero, and the probability that a locus carries the alternative or reference homozygote depends on allele frequencies and the inbreeding coefficient F, which ranges from zero in an F2 population to one in fully inbred doubled haploids. Crucially, the covariance between loci depends on the linkage disequilibrium coefficient D, which in turn is a function of the effective recombination rate between markers. Working through the algebra with allele frequencies of one half, the authors arrive at a strikingly compact result: the genotypic variance equals one plus F multiplied by a sum over chromosome pairs of twice the linkage disequilibrium coefficient times the product of the two marker effects.

This compact formula carries a subtle but important feature: linkage phase is captured by the signs of the marker effects themselves, with same signs indicating coupling and opposite signs indicating repulsion, rather than being embedded in the disequilibrium coefficient as in the classical formulation of Lynch and Walsh. The authors verified that their expression reduces to the classical result when F equals zero, and they checked the algebra with small-scale numerical examples. From the variance formula, the covariance between two populations follows from a simple identity: the variance of the sum of two populations’ genotypic values equals the sum of their individual variances plus twice their covariance. By computing the variance of the combined population using summed marker effects, and subtracting the two individual variances, the between-population covariance drops out analytically.

One practical obstacle is that the method nominally requires a linkage map to compute recombination rates between markers. Maps can be inaccurate, based on the wrong population, inflated by generations of random mating, or simply unavailable when only physical positions are known. The authors provide a workaround: for fully inbred populations and for F2 populations, the disequilibrium coefficient can be estimated directly from observed marker and haplotype frequencies as the absolute difference between the observed frequency of double-minor-allele homozygous haplotypes and the product of the individual allele frequencies. When phasing information is unavailable and the population is not fully inbred, published approximations from Ragsdale and Gravel can supply an unbiased estimate. This flexibility means the framework can be applied even when map resources are poor, a meaningful advantage for orphan crops and under-resourced breeding programs.

To validate the approach, the authors simulated four nested populations of one thousand individuals each, built so that each successive population segregated for only half as many quantitative trait loci as the one before. Populations were generated at three inbreeding levels, F2, F3, and doubled haploid, using the AlphaSimR simulation package with maize-like genome parameters: ten chromosomes of two Morgans each, carrying twenty quantitative trait loci apiece. Across one thousand independent simulation repeats, the analytical estimates tracked the true variances and covariances with remarkable fidelity. The correlation between observed and estimated variance exceeded 0.8 in every repeat, with a mean of 0.99, whether the known linkage map or the estimated disequilibrium coefficients were used. As theory predicted, variances and covariances in F2 populations were half those of doubled haploids, and F3 populations fell at seventy-five percent, confirming that the framework behaves correctly across the full range of inbreeding.

The real-world test came from a commercial maize breeding program at AgReliant Genetics. Four biparental populations of doubled haploid lines, each crossed to two of four tester inbreds, were evaluated in multi-environment field trials across four to ten locations in the U.S. Corn Belt over three years. Five traits were recorded: grain yield, grain moisture at harvest, plant height, ear height, and test weight. Marker effects for each population were estimated separately using ridge regression best linear unbiased prediction, and the authors discovered that the choice of shrinkage factor mattered enormously. When shrinkage was based on the raw number of markers, analytical variance estimates were consistently biased downward; when based on the number of fifty-centimorgan chromosome segments, they were biased upward. Using the effective number of independent tests, calculated from the eigenvalues of the marker correlation matrix following Li and Ji, yielded near-unbiased estimates that correlated strongly, above 0.97, with classical restricted maximum likelihood estimates.

The estimated covariances between maize populations revealed trait-specific and population-specific structure. For grain yield, covariances between population pairs ranged from negative values up to 9.52 square tonnes per hectare, with a mean of 1.67. The correlations between marker effects across populations were generally low to moderate, averaging 0.10 for yield, and the highest covariances appeared, unsurprisingly, between the same population tested with different testers. The lowest covariances occurred between populations sharing no parents. Notably, the correlation between the estimated genetic covariance and the marker-effect correlation was significant for most traits, reaching 0.77 for test weight, underscoring that between-population covariance reflects both relatedness and the genetic architecture of each trait. The authors also showed that their approach scales gracefully: where a joint multitrait analysis of fifty populations with five thousand markers would require estimating 250,000 marker effects and 1,275 covariance parameters simultaneously, their method decomposes the problem into fifty independent analyses.

The implications stretch well beyond the maize field. In breeding, heterogeneous genetic covariances could be folded into prediction models to improve accuracy when training and target populations differ, a chronic challenge in genomic selection. In quantitative trait locus mapping, the framework may allow genetically related populations to be pooled into larger discovery panels, boosting the power and resolution of gene detection. Beyond agriculture, the authors point to ecology and evolutionary genetics, where quantifying genetic covariance between subpopulations could illuminate hybrid zones, inform conservation of genetic diversity, and sharpen studies of population structure and adaptation. The method does carry limitations: it assumes a uniform inbreeding level within each population, and it requires marker effects, which are population-specific and depend on phenotypic data that may not yet exist for untested or uncreated populations.

What makes the study conceptually significant is that it fills a gap that has persisted since the founding of quantitative genetics. Covariance within populations has been understood since Fisher; covariance between traits and environments since the 1930s and 1950s. But the covariance between populations, the quantity that governs how much genetic information flows across breeding program boundaries, subspecies divides, and heterotic group walls, had no formal definition or estimator. By grounding between-population covariance in shared marker alleles and validating it against both simulated truth and commercial field data, Rebollo and Bernardo have given geneticists a new measurement tool that is analytically tractable, computationally scalable, and applicable at any level of inbreeding. As genomic datasets continue to accumulate across crops, livestock, and wild species, a framework for measuring how populations co-vary genetically may prove as foundational as the one Fisher built for measuring how relatives do.

Subject of Research: Estimation of genetic covariance between plant biparental populations using molecular marker effects

Article Title: Genetic covariance between plant biparental populations: concept and estimation

Article References: Rebollo, I., & Bernardo, R. (2026). Genetic covariance between plant biparental populations: concept and estimation. Heredity. https://doi.org/10.1038/s41437-026-00887-w

Image Credits: AI Generated

DOI: 10.1038/s41437-026-00887-w

Keywords: genetic covariance, quantitative genetics, plant breeding, biparental populations, marker effects, linkage disequilibrium, maize, genomic prediction, inbreeding coefficient, doubled haploids, RR-BLUP, Heredity

The research focused on the IGF1R gene, a central player in the growth hormone signaling pathway that governs skeletal development, muscle growth, and overall body conformation in mammals. As the receptor for insulin-like growth factor 1, IGF1R mediates the core biological functions of the IGF system, which is involved in cell proliferation, differentiation, and metabolism. The gene is widely expressed across tissues, including liver, muscle, bone, heart, lung, and brain, and plays critical roles during both embryonic and postnatal development. Because of this pivotal position in the growth axis, the researchers considered IGF1R a prime candidate for harboring genetic variants that influence economically important traits in livestock.

To investigate, the team selected 400 healthy, three-year-old female Qinghai Plateau yaks from a cooperative in Qilian County, Haibei Tibetan Autonomous Prefecture. Each animal was measured for five key growth indicators: body weight, oblique body length, body height, chest circumference, and cannon bone circumference. Blood samples were then collected from the jugular vein, anticoagulated, and stored at minus 80 degrees Celsius. Genomic DNA was extracted and its quality verified using spectrophotometry, with all samples exceeding 50 nanograms per microliter in concentration and maintaining an optical density ratio above 1.8, ensuring reliable material for downstream genetic analysis.

The genotyping approach relied on polymerase chain reaction amplification of IGF1R fragments, followed by Sanger sequencing. Primers were designed from the reference yak IGF1R sequence deposited in GenBank, and amplified products were verified by agarose gel electrophoresis before purification and sequencing. The yak IGF1R gene spans 305,476 base pairs and contains 21 exons and 20 introns. From the sequencing data, the team identified four SNPs: one, designated g.7383084C>T, located in exon 13, and three others, g.7383136G>A, g.7383137C>T, and g.7383177G>A, located in the adjacent intron 13. Sanger chromatograms clearly distinguished homozygous individuals, showing single fluorescent peaks, from heterozygotes, which displayed overlapping peaks corresponding to both nucleotides.

Statistical analysis of the population genetics revealed that all four loci were in complete Hardy–Weinberg equilibrium, indicating a relatively stable genetic state within the sampled herd. Three of the variants, g.7383084C>T, g.7383136G>A, and g.7383137C>T, exhibited moderate polymorphism, with polymorphism information content values between 0.25 and 0.5, suggesting sufficient allelic diversity to be useful for genotype-phenotype association studies. The fourth variant, g.7383177G>A, showed low polymorphism and limited variability. Genotype counts varied considerably across loci: for instance, at the exonic site, 182 animals carried the CC genotype, 182 carried CT, and 36 carried TT, while at the last intronic site, 382 of the 400 yaks carried the GG genotype and only 18 carried GA.

The association analysis, which accounted for age and environment as covariates in a fixed statistical model, produced striking results. All four SNP loci were significantly or highly significantly associated with body height and body length. At the exonic g.7383084C>T site, the variant was also linked to body weight and chest circumference, with CC and CT carriers showing significantly higher body weight than CT-adjacent patterns, and chest circumference increasing from CC to CT and TT animals, hinting at a possible positive dose-related effect of the T allele. At g.7383137C>T, heterozygous animals tended to excel in body height, chest circumference, and cannon bone circumference, while homozygous TT animals showed the greatest body length, a pattern that complicates simple additive interpretations.

Beyond individual loci, the researchers performed linkage disequilibrium and haplotype analysis using Haploview software. The four SNPs showed only weak linkage with one another, with pairwise D-prime values all below 0.8 and r-squared values below 0.33. From the combinations of alleles across the four sites, the team identified six haplotypes at frequencies above 3 percent, with Hap2 dominating at a frequency of 0.263. When the five common diplotype combinations, meaning paired haplotypes inherited from both parents, were tested against the growth traits, all five showed significant or highly significant associations with at least some measurements. One combination, H2H6, stood out as the optimal diplotype, associated with significantly higher body weight, withers height, body length, and chest circumference than most alternatives.

The biological plausibility of these associations is rooted in the molecular function of IGF1R. The receptor mediates IGF1-dependent signaling that directly regulates chondrocyte proliferation, longitudinal bone growth, and systemic skeletal development, all fundamental processes shaping body size. Notably, the exonic variant g.7383084C>T turned out to be a synonymous substitution: both alleles encode the amino acid aspartic acid, so the protein sequence is unchanged. Yet the authors caution against dismissing synonymous variants as functionally neutral. Such mutations can influence codon usage bias, messenger RNA stability and secondary structure, translation efficiency, co-translational protein folding, and exonic splicing regulatory elements, any of which could alter receptor production and, in turn, growth.

The three intronic variants are equally intriguing from a mechanistic standpoint. Two of them, g.7383136G>A and g.7383137C>T, sit just 9 and 10 base pairs downstream of exon 13, placing them within near-exon intronic regions that can participate in splice-site recognition and early spliceosomal assembly. Variants in these positions may perturb splice donor recognition, exon definition, or intronic splicing enhancers and silencers, potentially causing exon skipping, intron retention, or activation of cryptic splice sites. Such aberrant transcripts could change receptor isoform composition or expression levels. The researchers also note that the observed genotype patterns were not strictly linear for all traits, with heterozygote superiority appearing at some loci, which may reflect dominance effects, overdominance, or linkage with other functional variants in the local haplotype background.

The broader significance of the study extends beyond yak biology. IGF1R polymorphisms have been linked to growth, carcass, and meat quality traits across cattle, sheep, and pigs, but genetic effects often vary by breed, as seen when copy number variants of IGF1R associated with body size in Jinnan and Qinchuan cattle but not in Nanyang or Xianan cattle. The new yak data confirm that sequence variation in both exonic and flanking intronic regions of the gene contributes to phenotypic diversity across species and genetic backgrounds. With China’s yak population exceeding 15 million head, more than 95 percent of the global total, the practical implications are substantial. Incorporating these molecular markers into breeding programs could enhance selection efficiency, shorten breeding cycles, and support sustainable development of the yak industry, benefiting the plateau communities that depend on these remarkable animals for meat, milk, draft power, and cultural continuity.

Subject of Research: Associations between IGF1R gene polymorphisms and growth traits in Qinghai Plateau yaks

Article Title: Unveiling potential genetic markers: associations of IGF1R eon 13 and Intronic polymorphisms with growth traits in yak

Article References: Liu, X., Han, Y., Sun, Y., Gou, F., Chen, J., Jiang, W., & Zhao, Q. (2026). Unveiling potential genetic markers: associations of IGF1R eon 13 and Intronic polymorphisms with growth traits in yak. Archives Animal Breeding, 69(4), 553-561. https://doi.org/10.5194/aab-69-553-2026

Image Credits: AI Generated

DOI: 10.5194/aab-69-553-2026

Keywords: yak, IGF1R, SNP, growth traits, molecular markers, haplotype, marker-assisted selection, Qinghai Plateau, livestock genetics, linkage disequilibrium, synonymous mutation, animal breeding

Genotype phasing matters because most sequencing technologies read a person’s genome as an unordered mix of genetic variants. At any position where the two chromosomes differ, standard sequencing tells you that a variant is present but not which parent’s chromosome carries it. Reconstructing which variants travel together along each chromosome, producing what geneticists call haplotypes, is essential for understanding inheritance, tracing ancestry, identifying disease-associated regions, and interpreting the functional consequences of mutations. Errors in phasing, particularly so-called phase switches where the assignment flips incorrectly between adjacent variants, can distort downstream analyses and lead to mistaken conclusions about how genetic variation contributes to health and disease.

The dominant phasing tools in use today, including Eagle, Beagle, and Mendel Impute, achieve high accuracy by comparing an individual’s genotypes against large panels of previously phased reference genomes. These methods exploit linkage disequilibrium, the non-random association of variants that arises because stretches of DNA are inherited together across generations. The catch is that the approach depends on having a reference panel that closely matches the ancestry of the individual being phased. Populations that are underrepresented in these panels, particularly those with greater genetic diversity such as African populations, tend to receive less accurate phasing. The reference panels themselves are also enormous, demanding substantial computational resources and limiting scalability as sequencing efforts expand worldwide.

RefFree-Phaser takes a fundamentally different route. Rather than consulting a library of other people’s genomes, the model learns the statistical structure of genomic inheritance directly from sequence data. At its core sits BigBird, a transformer architecture designed to handle long sequences efficiently through sparse attention. Standard transformers compute attention between every pair of positions in a sequence, an operation whose cost grows quadratically with length, which makes them impractical for genomic data where relevant interactions can span millions of base pairs. BigBird’s sparse attention restricts each position to attending over a limited set of neighbors plus a set of global tokens, dramatically reducing computational expense while still allowing the model to capture long-range dependencies across the genome.

The technical pipeline behind RefFree-Phaser reflects careful adaptation of transformer machinery to genomic data. Input sequences are first tokenized and enriched with positional embeddings so the model understands where each variant sits along the chromosome. These representations then pass through a twelve-layer BigBird transformer, producing contextualized hidden states that encode information about each variant in light of its genomic surroundings. From these hidden states, the model derives logits, confidence scores, and binary haplotype predictions that align with the unphased genotypes and true labels used during training. To make the output unambiguous, the researchers adopted a lexicographical haplotype-pairing strategy in which all possible haplotype configurations are sorted and systematically tagged according to their input genotypes, ensuring a consistent and reproducible mapping between predictions and the underlying data.

The evaluation spanned two cohorts, one drawn from a subset of the 1000 Genomes Project and the other from the Omni2.5 million common-variants dataset. Across diverse populations, RefFree-Phaser achieved average phasing accuracies of 92.60 percent on the 1000 Genomes data and 91.56 percent on the Omni2.5 million common variants, figures that place it in the same territory as the established reference-based methods it was compared against. Performance was strongest for individuals of European ancestry and other superpopulations, while accuracy dipped slightly for African individuals. The authors attribute this gap to higher phase-switch rates, greater genetic diversity, and limited representation in the training data, a pattern that mirrors known limitations of the field rather than a flaw unique to this model.

One of the most intriguing findings is that the model does more than simply match the accuracy of classical tools. RefFree-Phaser captured genomic behaviors that conventional methods do not explicitly report, most notably phase switches at heterozygous sites, which the model flags with distinguishable confidence scores. In principle, this means the system not only phases genotypes but also communicates how certain it is at each position, giving researchers a quantitative signal for identifying regions where phasing is unreliable. Such built-in uncertainty estimates could prove valuable for downstream applications, allowing analysts to weight or filter genomic regions according to the confidence attached to their phase assignments.

Perhaps the most striking evidence that the transformer has genuinely learned genomic biology comes from an analysis of its attention patterns. The researchers found that the model’s attention mechanism is driven primarily by genetic relevance, and that the positions receiving the maximum cumulative attention align precisely with linkage disequilibrium markers and with true haplotype crossover sites, the genomic locations where recombination has shuffled genetic material between chromosomes. In other words, without being explicitly taught about recombination, the model appears to have discovered where crossovers occur, validating the contextualized hidden states as a robust feature set that encodes real biological structure rather than superficial statistical patterns.

The implications extend beyond a single benchmark result. A phasing method that requires no external reference panel could be transformative for genomic medicine in populations that reference-based tools serve poorly, and for large-scale sequencing projects where the computational burden of maintaining and querying massive panels becomes prohibitive. It also fits into a broader trend of deep learning models, from protein structure prediction to regulatory element annotation, demonstrating that neural networks can internalize complex biological regularities when given sufficient data and appropriate architectures. The sparse-attention design of BigBird, originally motivated by the need to process long documents, turns out to be well suited to the analogous challenge of long genomic sequences.

Challenges remain before such tools reach routine use. The reduced accuracy for African ancestry individuals underscores a persistent problem in genomics: training data itself carries the biases of historical sampling, and a reference-free model is only as representative as the data it learns from. The study, supported by UTHealth startup funds and a National Human Genome Research Institute grant, was led by Muhammad Nadeem Cheema, with Anam Nazir and Arif Harmanci as co-authors, all based in the Department of Health Data Science and Artificial Intelligence at the D. Bradley McWilliams School of Biomedical Informatics. Their work, published open access on 29 August 2026, demonstrates that genotype phasing can indeed be performed without external reference panels using a transformer-based model that captures long-range genomic dependencies. As training datasets grow more diverse and transformer architectures continue to improve, reference-free approaches like this one may reshape how the field reconstructs the two intertwined stories that every genome tells.

Subject of Research: Reference-free genotype phasing using a BigBird transformer deep learning model

Article Title: RefFree-Phaser: a reference-free transformer-based framework for genotype phasing

Article References: Cheema, M. N., Nazir, A., & Harmanci, A. (2026). RefFree-Phaser: a reference-free transformer-based framework for genotype phasing. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06572-3

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06572-3

Keywords: genotype phasing, haplotype inference, BigBird transformer, deep learning, genomics, linkage disequilibrium, 1000 Genomes Project, reference-free methods, sparse attention, recombination, BMC Bioinformatics, population genetics

The researchers focused on S3011, an interspecific population derived from a cross between cultivated annual sorghum, Sorghum bicolor, and its wild perennial relative, Sorghum halepense, commonly known as Johnson grass. After hybridization, diploid progeny were recovered, and a structured F4 panel was developed for field evaluation. This material is scientifically valuable because perenniality in sorghum is thought to have been lost during domestication, and reintroducing it from the wild relative could, in principle, allow grain crops that regrow year after year without replanting. Rhizomes, the underground stems from which new shoots emerge, are a central component of that perennial persistence.

The team phenotyped the population under field conditions for two related but distinct traits: rhizome number and overwinter regrowth. Accession-level best linear unbiased predictors, or BLUPs, were computed to reduce noise from environmental and spatial variation, and these adjusted values were then used for genome-wide association analyses. Genotyping-by-sequencing yielded 43,317 single-nucleotide polymorphisms across the genome, providing a dense enough marker framework to scan for associations between genetic variants and the perenniality traits.

The headline statistical result is a study in transparency: no association exceeded the Bonferroni-adjusted genome-wide significance threshold of P at or below 1.15 times ten to the minus six, and none remained significant after Benjamini-Hochberg false discovery rate correction at a q-value below 0.05. In other words, by the strictest standards, the genome-wide association study produced no confirmed quantitative trait loci. Rather than burying this outcome, the authors used a predefined suggestive threshold of P below 0.0001, corresponding to a negative log10 P-value greater than 4, to identify 13 SNPs, which they consolidated into seven spatially distinct exploratory signals for rhizome number.

What elevates the study beyond a standard suggestive-threshold exercise is the battery of complementary analyses applied to each candidate region. The researchers examined high-low allele-dosage contrasts, asking whether plants carrying more copies of the associated allele produced more rhizomes than those carrying fewer. They characterized local linkage disequilibrium to understand how far each signal extended across the chromosome. They checked historical positional correspondence, comparing their regions with genomic intervals previously implicated in perenniality-related traits. They annotated positional candidate genes within each region. Finally, they ran a descriptive marker-based genome scan as an independent analytical method on the same dataset.

The strongest allele-dosage differentiation occurred at a signal the authors designate chromosome 1 cluster 2, marked by SNP S1_63359080, with a contrast value of plus 0.3364 that placed it at the 100.0th chromosome-relative percentile. A neighboring signal, chromosome 1 cluster 1 at SNP S1_62967991, showed a contrast of plus 0.3002 at the 99.3rd percentile. These two adjacent regions on chromosome 1 emerged as the most compelling of the seven exploratory signals, combining suggestive GWAS evidence with pronounced allele-dosage differentiation. The complementary marker scan showed within-dataset cross-method concordance for selected rhizome-number signals, including a region on chromosome 3, suggesting that the association patterns were not artifacts of a single analytical pipeline.

Overwinter regrowth told a different story. The trait produced three exact marker-level matches between the GWAS and the marker scan, but it lacked the localized multi-marker regional structure observed for rhizome number. This contrast is biologically meaningful: rhizome development appears to leave a more coherent genomic footprint, with several nearby markers reinforcing one another, whereas overwinter regrowth may be controlled by a more diffuse set of loci, by stronger environmental influences, or by both. The authors emphasize that rhizome development and overwinter regrowth, though both components of perennial persistence, may differ genetically and environmentally, and their data support treating them as separable breeding targets rather than a single trait.

The authors are unusually explicit about the limits of their concordance evidence. Because both the GWAS and the marker scan used the same S3011 dataset, agreement between them represents cross-method concordance rather than independent validation. Given the modest sample size, the low heritability of the traits, and the absence of associations surviving multiple-testing correction, all identified signals remain exploratory. The paper states plainly that the results do not establish confirmed QTL or causal loci. This kind of calibrated language is rare and welcome in a field where suggestive associations are sometimes oversold, and it sets a clear agenda: the seven rhizome-number regions and the regrowth markers now need testing in independent populations and environments before any breeding utility can be claimed.

For the perennial agriculture movement, the study is best understood as a prioritization tool. By ranking candidate regions according to allele-dosage differentiation, local linkage disequilibrium structure, historical correspondence with prior QTL, and candidate-gene content, the researchers have generated a short list of genomic neighborhoods worth pursuing. The chromosome 1 clusters stand out, the chromosome 3 region gains support from cross-method agreement, and the candidate-gene annotations provide starting points for functional work. Validation in larger, independent populations, ideally with higher heritability through multi-environment testing, will determine whether any of these exploratory signals can be converted into markers useful for breeding perennial sorghum.

The broader significance lies in the honest template the study offers for complex-trait genomics in orphan and emerging crop systems. Perenniality is a polygenic, environmentally sensitive trait, and the reality is that a single F4 panel will rarely deliver definitive loci. By combining a conventional GWAS with allele-dosage contrasts, linkage disequilibrium profiling, historical QTL comparison, candidate-gene annotation, and a parallel genome scan, and by reporting every negative result, Nabukalu and colleagues have shown how exploratory genomic signals can be characterized responsibly. The work was supported by the Malone Family Land Preservation Foundation and the Perennial Agriculture Project, a joint initiative of The Land Institute and the foundation. If the prioritized regions hold up in validation, the study will be remembered as an early step toward grains that come back on their own each spring; if they do not, the framework for evaluating them will remain a model worth copying.

Subject of Research: Genetic architecture of rhizome development and overwinter regrowth in diploid perennial sorghum

Article Title: Genetic architecture of rhizome development in diploid perennial sorghum reveals exploratory genomic signals with heterogeneous complementary support

Article References: Nabukalu, P., Kong, W., Cox, S., Reeves, J., & Paterson, A. H. (2026). Genetic architecture of rhizome development in diploid perennial sorghum reveals exploratory genomic signals with heterogeneous complementary support. BMC Plant Biology. https://doi.org/10.1186/s12870-026-09930-8

Image Credits: AI Generated

DOI: 10.1186/s12870-026-09930-8

Keywords: perennial sorghum, rhizome development, overwinter regrowth, genome-wide association study, Sorghum halepense, allele dosage, linkage disequilibrium, quantitative trait loci, plant genetics, perennial agriculture, SNP markers, interspecific hybridization

The core innovation of KNOT lies in its use of knockoffs, a concept borrowed from modern high-dimensional statistics. A knockoff is a deliberately constructed decoy variable that mimics the correlation structure of the original genetic variants but has no genuine relationship with the phenotype being studied. By presenting a neural network with both real variants and their knockoff counterparts, researchers can ask a pointed question: does the model treat a real variant as more important than its decoy? If a variant consistently outranks its knockoff across many comparisons, that variant earns a place on the candidate list. Crucially, the proportion of decoys that slip through provides a direct estimate of how many false positives are contaminating the results. This is the essence of false discovery rate control, a guarantee that conventional deep-learning pipelines for genetic association simply do not offer. KNOT wraps its neural network inside this knockoff framework, producing variable selection that is both flexible enough to capture complex genetic patterns and disciplined enough to quantify its own error rate.

Stabilizing the network’s stochastic behavior is a second pillar of the method. Neural networks initialized with different random seeds can produce different importance rankings, which is unacceptable when the goal is to nominate specific variants for biological follow-up. KNOT addresses this instability through contrastive learning, a training strategy that teaches the model to distinguish between related and unrelated samples. In family-based association studies, where parents and siblings share predictable fractions of their genomes, modeling sample relatedness is not optional; it is the difference between a signal that reflects genuine inherited risk and one that merely reflects family structure. By learning representations that account for these kinship relationships, KNOT reduces the variance of its selections and produces rankings that remain consistent across repeated runs, a property the authors identify as essential for trustworthy variable selection in genetic data.

Attention mechanisms supply the third technical ingredient. Within any genomic region, nearby variants travel together through generations as inherited blocks, a phenomenon known as linkage disequilibrium. An attention-based architecture allows the network to weigh the contributions of all variants in a block simultaneously, learning which positions within a correlated cluster carry the strongest association with disease. This is a fundamentally different approach from the marginal testing of classical GWAS, which treats each variant in isolation and can attribute a signal to the wrong member of a correlated group. By explicitly modeling the correlation structure rather than ignoring it, KNOT aims to sharpen the localization of risk variants, bringing the nominated positions closer to the variants that actually matter biologically.

Epistasis, the interaction between variants at different genomic locations, represents one of the most tantalizing and least explored dimensions of genetic risk. Two variants may be harmless individually yet consequential in combination, a pattern that single-marker tests are structurally blind to. Exhaustive testing of all pairwise interactions across millions of variants is computationally prohibitive and statistically punishing, because the sheer number of tests inflates the burden of multiple-comparison correction. KNOT incorporates a permutation test to detect these epistatic interactions, using label permutation to establish a null distribution against which observed interaction effects can be judged. The permutation approach preserves the correlation structure of the data while destroying any genuine relationship between genotype and phenotype, providing a principled benchmark for deciding when an interaction is real.

The choice of a family-based study design is itself significant. Family-based association studies compare the alleles transmitted from parents to affected children against those not transmitted, a design that is inherently robust to confounding from population stratification, the spurious associations that arise when cases and controls differ in ancestry. This robustness comes at a cost: family designs typically have fewer effective samples and more complex correlation patterns than population-based case-control cohorts, which is precisely why deep-learning methods that ignore relatedness have struggled in this setting. KNOT was engineered from the outset for this environment, with its contrastive learning component explicitly modeling the relatedness that family data impose. The method therefore occupies a deliberately chosen niche, applying modern representation learning where its statistical hazards are most acute and its potential payoff is greatest.

To demonstrate the method’s practical value, Fu and Yang applied KNOT to two autism spectrum disorder family cohorts. Autism spectrum disorder is a compelling test case for any new genetic discovery method. Its heritability is high, yet much of that heritability remains unexplained by the variants identified to date, a puzzle often described as the missing heritability problem. Epistatic interactions are frequently invoked as one plausible reservoir of this unexplained risk, making a method that can jointly identify risk variants and their interactions especially relevant. In these applications, KNOT outperformed conventional methods, identifying more known and putative risk loci as well as additional epistatic interactions associated with autism. The recovery of known loci serves as a sanity check that the method is finding real signals, while the putative novel loci and interactions represent hypotheses for the field to pursue.

The broader significance of this work extends beyond autism genetics. The challenge that KNOT addresses, namely how to extract interpretable and statistically controlled discoveries from a flexible neural network applied to correlated high-dimensional data, is a defining problem of the current era of machine learning in the life sciences. Knockoff-based frameworks offer a general template: the neural network provides the representational power to model complex dependencies, while the knockoff construction provides the inferential scaffolding to keep the resulting claims honest. If this template generalizes, it could reshape how deep learning is deployed across association studies of other complex traits, from cardiovascular disease to psychiatric conditions, wherever correlated predictors and limited sample sizes conspire to produce fragile discoveries.

There are, of course, caveats that temper the enthusiasm. Knockoff construction depends on an accurate model of the correlation structure among variables, and genomic linkage patterns vary across populations; imperfect knockoffs can degrade the precision of the false discovery rate guarantee. Neural network training remains computationally demanding, and the permutation testing for epistasis adds further cost. The authors report their work from the Department of Biostatistics at City University of Hong Kong, supported by the Research Grants Council of Hong Kong Early Career Scheme, and the method’s performance will ultimately be judged by how well its discoveries replicate in independent cohorts, the ultimate arbiter of any genetic association claim. The study, published open access on 24 September 2026 in Genome Biology under a Creative Commons Attribution 4.0 license, invites exactly that scrutiny.

What makes KNOT a noteworthy development is not any single component but the synthesis. Contrastive learning tames the randomness of neural training, attention mechanisms respect the correlation architecture of the genome, knockoffs impose a quantifiable error budget on the selections, and permutation testing opens a window onto variant interactions that conventional pipelines cannot see. Applied to two autism family cohorts, this combination surfaced more known and putative risk loci and more epistatic interactions than established approaches. For a field that has spent years wrestling with the trade-off between the flexibility of deep learning and the rigor of classical statistics, KNOT offers a persuasive demonstration that the two need not be enemies. The variants and interactions it nominates now join the long queue of candidates awaiting biological validation, but the statistical machinery that nominated them marks a meaningful step toward deep-learning-driven genetics that researchers can actually trust.

Subject of Research: A knockoff-augmented deep learning method for identifying genetic risk variants and epistatic interactions in family-based genome-wide association studies of autism

Article Title: KNOT: a knockoff-augmented neural network for identifying risk variants and epistatic interactions in family-based association studies

Article References: Fu, Y., & Yang, Y. (2026). KNOT: a knockoff-augmented neural network for identifying risk variants and epistatic interactions in family-based association studies. Genome Biology. https://doi.org/10.1186/s13059-026-04285-7

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04285-7

Keywords: GWAS, deep learning, neural networks, knockoff statistics, epistasis, autism, family-based design, false discovery rate, linkage disequilibrium, contrastive learning, attention mechanisms, Genome Biology

The story begins with two breakthroughs made by the United States Department of Agriculture in the 1940s. The first was the discovery of cytoplasmic male sterility, or CMS, a trait that renders plants unable to produce viable pollen, allowing breeders to use them as female parents without laborious manual emasculation. The second was the genetic control of monogermity, a single recessive allele that causes each fruit to carry just one seed, eliminating the need for painstaking thinning of crowded seedlings. Together, these traits made large-scale hybrid seed production feasible in a crop where self-fertilization triggers severe inbreeding depression, sapping vigor and fertility. To sidestep that problem, breeders in the late 1950s developed the three-way hybrid scheme: a male sterile line is crossed with a maintainer line to produce vigorous hybrid female seed parents, which are then crossed with fertile male pollinators to yield the commercial varieties farmers plant.

This architecture creates a fundamental asymmetry. Two generations separate the female parents from the evaluated hybrids, while only one separates the males, and the two pools are managed under partly different constraints. The female side must carry CMS, monogermity, and adequate seed production, while the male side is selected mainly as the pollinator component, judged largely on root and sugar traits. Over sixty years, breeders have sought to maximize general combining ability between the pools, a process expected to fix different alleles in each pool, much as occurred in maize after its transition to hybrid breeding. Yet the genomic consequences of this scheme had never been systematically documented in sugar beet until now.

To fill that gap, the team grew and genotyped 11,099 plants between 2015 and 2023, representing 299 female accessions, 605 male pollinators, and 381 potential commercial three-way hybrids. Each plant was screened with more than 10,000 single-nucleotide polymorphisms positioned across the nine chromosomes of the sugar beet reference genome. Because breeding accessions can be genetically heterogeneous, the researchers developed a likelihood-based method to distill a consensus genotype for each accession from its progeny, carefully filtering out labeling errors and pollen contamination. The resulting dataset, spanning breeding years from 1998 to 2018, allowed the team to reconstruct the genetic history of an active commercial program with a resolution rarely possible outside long-term maize experiments.

The genetic structure analysis delivered an unambiguous verdict. The first principal component, explaining just over twenty percent of the variation, cleanly separated the male pool from the female hybrid pool, with commercial varieties occupying an intermediate, admixed position, exactly as expected for the offspring of two divergent parents. Within the female pool, the male sterile lines and their maintainers were also clearly differentiated. More striking was the temporal signal: the second principal component arranged male accessions along a gradient from oldest to most recent, and admixture analysis subdivided the males into three distinct breeding eras spanning 1998 to 2004, 2005 to 2011, and 2012 to 2018. No comparable temporal structure appeared in the female pools, reflecting their slower, more constrained breeding cycle.

Diversity within the male pool declined steadily over the twenty-one years, with modified Rogers’ distance falling from 0.35 in the earliest era to 0.33 in the most recent, a small but statistically significant erosion. Meanwhile, differentiation between males and the female hybrid reference pool, measured by Fst, crept upward from 0.25 to 0.27, and the proportion of highly differentiated markers with Fst above 0.5 rose from 19.3 percent to 22.7 percent, a relative increase of eighteen percent. The increase was not uniform across the genome: chromosomes 3, 4, 8, and 9 showed the strongest divergence, and on chromosome 8 the mean Fst between the first and third eras jumped by twenty-two percent, concentrated in regions that were already highly differentiated. Intriguingly, the loss of within-pool diversity and the gain of between-pool differentiation were tightly correlated, suggesting that selection in the male pool traded internal diversity for complementarity with the female side.

Linkage disequilibrium, the non-random association of alleles along chromosomes, told a complementary story. Short-range LD was significantly lower in females than in males, and within males it declined markedly over time, with the mean correlation between marker pairs dropping by about twenty-five percent across the study period. Older male lines from 1998 to 2004 carried long stretches of correlated variation, with elevated LD extending over tens of megabases on several chromosomes, while recent lines showed much faster decay. The researchers attribute this to breeding practices that increase allele shuffling, such as recombining diverse material before selection. When the team grouped correlated markers into haploblocks, they found that males carried fewer effective haplotypes per block than either female subgroup, another sign of consolidation within the pollinator pool.

Selection scans, combining a differentiation-based Fst outlier test with a principal-component-based method, pinpointed 174 outlier SNPs that distinguish the parental pools, most of which clustered into seven large haploblocks on five chromosomes. Two of these regions were immediately recognizable. On chromosome 3, a 1.1 megabase signal encompassed the Rz1 locus, the dominant rhizomania-resistance gene introduced from Holly Sugar germplasm after 1984 and now deployed in most commercial hybrids. Far more surprising was the signal on chromosome 2 spanning the bolting locus B, which contains BvBTC1, the major gene controlling the annual versus biennial growth habit. Although the biennial allele is considered fixed in domesticated sugar beet, the team uncovered fifteen distinct haplotypes at this locus, with one haplotype carried by ninety-three percent of males but fewer than a quarter of female lines, and another showing the reverse pattern, a haplotypic split between the sexes that likely reflects multiple founder contributions rather than new functional variation.

Temporal analysis of the male pool added a third layer of insight. Of 701 SNPs showing significant allele frequency shifts across eras, only eighteen changed consistently in every interval, while 329 shifted gradually between the first and third periods and hundreds more showed abrupt, non-monotonic swings more consistent with genetic drift than with directional selection. One haploblock on chromosome 4 stood out as a candidate target of persistent selection: its most frequent haplotype climbed from thirty-seven percent of males in the first era to eighty-five percent in the third, while its Fst against females rose from 0.69 to 0.76. At the beet cyst nematode tolerance locus on chromosome 5, introgressed from wild relatives in the early 2000s, no individual SNP showed a selection signature, yet a male-specific tolerance haplotype surged from nine percent of males in the earliest era to forty-three percent in the most recent, demonstrating that haplotype-level analysis can detect recent introgressions that single-marker scans miss.

The authors are careful to note the limits of their study. The SNP array introduces ascertainment bias, the female pool lacked the temporal resolution needed for parallel time-series analysis, and genomic differentiation alone cannot prove functional complementarity, since the work did not directly link these patterns to hybrid performance or combining ability. Drift, founder effects, and trait introgression may all have contributed alongside directional selection. Even so, the study establishes that sixty years of three-way hybrid breeding have produced measurable genomic divergence between sugar beet’s parental pools, and that the last twenty-one years have deepened that divide while slowly depleting within-pool diversity. The seven haploblocks identified, harboring genes tied to flowering, disease resistance, and stress response, now provide a genomic roadmap for breeders and researchers seeking to understand how heterosis is built, and whether the haplotypic differences accumulated over two decades of selection translate into the agronomic performance that sugar beet’s continued success depends on.

Subject of Research: Genomic consequences of three-way hybrid breeding on male and female gene pools in sugar beet

Article Title: Twenty-one years of three-way hybrid breeding shaped diversity and complementarity of male and female gene pools in sugar beet

Article References: Desprez, A., Henry, K., Desprez, B., Devaux, P., Charcosset, A., Tenaillon, M. I., & Moreau, L. (2026). Twenty-one years of three-way hybrid breeding shaped diversity and complementarity of male and female gene pools in sugar beet. Theoretical and Applied Genetics, 139(10), Article 271. https://doi.org/10.1007/s00122-026-05366-8

Image Credits: AI Generated

DOI: 10.1007/s00122-026-05366-8

Keywords: sugar beet, hybrid breeding, heterosis, cytoplasmic male sterility, genetic diversity, linkage disequilibrium, haploblocks, Fst differentiation, BvBTC1, rhizomania resistance, beet cyst nematode tolerance, plant breeding