The statistics behind Scilytics software.

How Scilytics chooses its statistical methods, checks them and reports results.

Methods

Scilytics software uses published methods that are widely used in plant and animal breeding. Each product lists the method and software at each stage, for CacaoIQ in its technical notes.

  • Mixed models for field trials

    Trials are analyzed with linear mixed models that account for rows, columns and blocks, and give a BLUP for each entry.

  • Genomic relationships

    A genomic relationship matrix built from SNP markers (VanRaden 2008) measures how closely two individuals are related.

  • Association mapping

    Mixed-model GWAS corrects for relatedness with the kinship matrix (Kang et al. 2010) and tests each marker against a genome-wide threshold.

  • QTL mapping in full-sib families

    Interval mapping for outcrossed full-sib families estimates linkage phase and segregation pattern (Gazaffi et al. 2014), with a permutation threshold.

Validation

Every analysis is validated in two ways. First, it follows established statistical methods for breeding. Second, we reproduce published analyses on publicly available data and compare our results with the reported findings. When results differ, we investigate the underlying methods and assumptions to understand why. This can show us where to improve our software, clarify a difference in method, or reveal an inconsistency in the published results. Every comparison is documented, and each product's pages list the ones released so far.

Where a method has no public data to check it against yet, the product's pages say so.

CacaoIQ validation results

Data checks and reporting rules

  • Data checks

    Duplicates, outliers and impossible values are flagged for you to review. Poorly called samples and off-types are removed from genotype data and listed. Nothing is dropped without being reported.

  • Candidate genes

    Candidate genes are listed only at loci above the genome-wide threshold: Bonferroni for GWAS, the permutation threshold for QTL mapping.

  • Numbers and text

    The statistical models produce every number. A language model may explain a method or suggest column roles, but it never computes a result.

References

The papers behind these methods. Each product cites its own data, genome and software, for CacaoIQ in its technical notes.

  1. VanRaden PM (2008). Efficient methods to compute genomic predictions. Journal of Dairy Science 91: 4414-4423. doi:10.3168/jds.2007-0980
  2. Kang HM, Sul JH, Service SK, et al. (2010). Variance component model to account for sample structure in genome-wide association studies. Nature Genetics 42: 348-354. doi:10.1038/ng.548
  3. Gazaffi R, Margarido GRA, Pastina MM, Mollinari M, Garcia AAF (2014). A model for quantitative trait loci mapping, linkage phase, and segregation pattern estimation for a full-sib progeny. Tree Genetics & Genomes 10: 791-801. doi:10.1007/s11295-013-0664-2

Try CacaoIQ on a trial you know well.

The evaluation runs on your computer.

Request an evaluation