ارزیابی عملکرد ماشین بردار پشتیبان با کرنل های مختلف در تجزیه ژنومی در سطوح مختلف واریانس غالبیت

هدف از پژوهش حاضر، بررسی و مقایسه صحت پیش بینی ژنومی روش ماشین بردار پشتیبان (svm) بر اساس توابع کرنل مختلف شامل خطی (svm-lin)، شعاعی (svm-rad)، چند جمله ای (svm-pol) و حلقوی (svm-sig)، و روش gblup در مدل های کنش ژنی صرفاً افزایشی و افزایشی-انحراف غالبیت با در نظر گرفتن سطوح مختلف واریانس غالبیت بود. بدین منظور، ژنومی حاوی شش کروموزوم و به طول 600 سانتی مورگان شبیه سازی شد. روی هر کروموزوم، 1000 نشانگر چندشکلی‌ تک نوکلئوتیدی (snp) با فواصل یکسان و 100 جایگاه صفت کمّی (qtl) به طور تصادفی در نظر گرفته شد. واریانس فنوتیپی و وراثت پذیری به ترتیب برابر با 1 و 0.4 در نظر گرفته شد. واریانس انحراف غالبیت برابر با 0.10، 0.15، 0.20، 0.25، 0.30 و 0.35 در نظر گرفته شد. صحت پیش بینی به عنوان ضریب همبستگی پیرسون بین ارزش ژنتیکی واقعی (tgv) یا ارزش اصلاحی واقعی(tbv) و ارزش ژنتیکی ژنومی (gegv) یا ارزش اصلاحی ژنومی (gebv) تعریف شد. روش مرسوم gblup در تمام سناریوهای مختلف واریانس غالبیت، صحت پیش بینی gebv و gegv بالاتری را نشان داد. در بین رویکردهای مختلف svm، در مدل صرفاً افزایشی و افزایشی- انحراف غالبیت بر اساس صحت پیش بینی gegv، رویکردهای svm-rad و svm-sig به ترتیب بالاترین عملکرد را نشان دادند. بر اساس صحت پیش بینی gebv، با افزایش واریانس غالبیت، این برتری به شدت کاهش یافت، به ‌ طوری که در واریانس غالبیت بیشتر از 0.30، رویکردهای svm-lin و svm-sig به ترتیب صحت پیش بینی gebv اندکی بالاتر و برابر با svm-rad نشان دادند. به طورکلی، در برازش فنوتیپ روی نشانگرها با روش ناپارامتری svm، استفاده از تابع کرنل شعاعی در مدل پیشنهاد می شود.

evaluating the performance of support vector machine with different kernels in genomic analysis at different levels of dominance variance

introduction: predicting quantitative traits is a fundamental aspect of plant and animal breeding. genomic selection is a precise and efficient approach that estimates genetic merit using high-density single nucleotide polymorphisms (snps). however, most genomic selection procedures primarily focus on additive effects to calculate the genomic estimated breeding value (gebv) for selection candidates. nonetheless, incorporating non-additive effects offers several advantages: (i) it enhances the accuracy of gebv predictions and subsequent selection responses, (ii) it facilitates optimized mate allocation among selection candidates, and (iii) it enables improved utilization of non-additive genetic variation through tailored crossbreeding or purebred breeding strategies. one of the most challenging factors affecting the accuracy of genomic evaluation is the selection of an appropriate statistical method to estimate marker effects with high accuracy. the most common parametric methods for genomic evaluation are the genomic best linear unbiased predictor (gblup) and bayesian methods, which use the co-variance structure between individuals and regression of phenotype on markers to predict the genetic values ​​of individuals, respectively. however, in recent years, non-parametric methods of machine learning have been widely used for genomic evaluation in animal and plant breeding programs. this study aimed to evaluate and compare the accuracy of genomic predictions using gblup and support vector machine (svm) methods. the svm models employed various kernel functions, including linear (svm-lin), gaussian radial (svm-rad), polynomial (svm-pol), and cyclic (svm-sig). both purely additive and additive + dominance deviation gene action models were considered under varying levels of dominance variance.materials and methods: a simulated genome comprising six chromosomes with a total length of 600 cm was used for this study. each chromosome contained 1,000 evenly spaced snps and 100 randomly distributed quantitative trait loci (qtls). phenotypic variance ( ) and narrow-sense heritability ( ) were set to 1 and 0.4, respectively. dominance variance ( ) levels were evaluated at 0.10, 0.15, 0.20, 0.25, 0.30, and 0.35. prediction accuracy was calculated as the pearson correlation coefficient between the true genetic value (tgv) or true breeding value (tbv) and their respective genomic estimates (gegv or gebv). the correlations were represented as  and , respectively. in addition, the practical significance of differences in prediction accuracy among the studied statistical methods was assessed using cohen’s d effect size during 100 replicates.results and discussion: the conventional gblup method consistently exhibited higher prediction accuracy for both gebv and gegv across all scenarios. among the svm approaches, the svm-rad and svm-sig kernels showed superior performance in predicting gegv under both purely additive and additive + dominance deviation models. however, for gebv prediction, their performance declined with increasing dominance variance. when dominance variance exceeded 0.30, svm-lin and svm-sig demonstrated prediction accuracy comparable to or slightly better than svm-rad. the differences in prediction accuracy between gblup and svm-rad were minimal (d=0.218) in the purely additive model but reached their peak (d=0.492 and d=0.404) in the additive + dominance deviation model at the highest dominance variance ( ). this disparity occurred because, as dominance variance increased, the gblup method exhibited slightly greater changes in accuracy compared to the svm-rad method. furthermore, with higher dominance variance, the difference in prediction accuracy for gebv between the svm methods and both gblup and svm-rad substantially decreased. for example, in the purely additive model ( ), cohen’s d was 2.608 and 2.336, respectively, while in the additive + dominance deviation model ( ), d dropped to 0.309 and 0.189, respectively. using the additive + dominance deviation model significantly improved gebv prediction accuracy, particularly when dominance variance contributed substantially to phenotypic variance. this improvement is due to dominance deviation, which arises from interactions between alleles at a locus. while additive effects are represented as breeding values, which partially incorporate dominance effects, genomic evaluations that explicitly consider dominance effects can further enhance gebv accuracy.conclusions: the choice of kernel function in svm models plays a pivotal role in the accuracy of gebv and gegv predictions. overall, when applying the nonparametric svm method to fit markers to phenotypes, the gaussian radial kernel function is recommended for optimal performance. however, as the dominance variance increased, the performance of svm-lin and svm-sig methods improved significantly and the performance gap with gblup and svm-rad decreased. this indicated the potential capacity of svm in investigating non-additive effects, especially in situations where the contribution of dominance in explaining phenotypic variance increases.