Breeding beyond genomics

From the Statistical point of view, I tend to think Quantitative Genetics and related fields as domains where two main “pillars” cohabitate: Inference Prediction. Even if same tool, for example penalized linear models, can be used both tasks inference prediction may reinforce each other, they are distinct concepts. It is interesting observe how these pillars have reacted big data, that large p small n paradigm. While studies target tried (unsuccessfully) protect against false positives, practitioners embraced new era with joy. Why so? Very simple: Prediction falsifiable via cross-validation, whereas validation not straightforward, an increase in variables easily leads confounding. Most relevant distributional properties depend on knowing actual, “true” model. Both are, however, encountering serious problems. First, consider “Inference”. For many years, breeding involved a few parameters or very carefully chosen say including maternal effects. Today, literature flooded reports genome wide association (GWAS) signals selective footprints. In standard GWAS, when markers individually estimated without penalization, issue controlling positive rates. Identifying sweeps also tricky, numerous statistics coexist, pinpointing different regions. Further, significance well defined this task. Do get me wrong, am responsible some GWAS sweep studies. Large scale unrelated individuals from populations effective size useful. Understanding patterns DNA variability based solid theory. most livestock studies, though, one should take results caution been replicated independent turn has blessed multidimensionality. As long penalization cross-validation properly employed, having more desirable than only few. Success number predictors increased astounding plant fields, genetic progress accelerated since application genomic selection. This possible, insist, because and, therefore, pragmatism dominates. Of note, model predict even causative mutations so better performing closest “biological causality.” The Achilles heel interpretability. machines “black boxes,” although degree “opacity” varies. GBLUP allows at least recovering marginal marker effects, convolutional neural networks do not. all, interpretability non-negligible regarding communication methods industry society. available, yet perform similarly. Have we reached “methodological” plateau? skills high demand worldwide, Breeding mature field scientific advances seem incremental. disciplines, animal breeders’ population rather inbred, will likely by looking inspiration outside our own science. Where challenges breeding, then? optimistic. see exciting prospects several areas let just mention Phenomics, automatic measurement phenotypes, far attractive challenge, my opinion. Highly unstructured, massive heterogeneous datasets now cheaply produced sensors. New opportunities exist developing algorithms transform raw data into meaningful phenotypes implementing programmes dimensional data. Among analysing individual group behaviour via, say, video recording problem. Impact topic utmost interest terms research, industry, evolutionary experiments provide unique biological knowledge. second domain foresee discoveries, once longitudinal phenomic sets available. Animal genomes highly resilient but responsive; pressure result (slightly) allele frequency changes. Besides, response selection almost never exhausted. intriguing observation highlights relevance physiological mechanisms activated concerted action stages. Finally, domestication terrestrial species rare phenomenon human history. Only handful domesticated, behavioural reproductive constraints. scenario completely aquaculture, dozens recently started grown captivity, process. aquaculture general advanced technologically farming poses practical methodological challenges. But extended broadly, insects feeding. There uncharted territories curious breeder. finish thanking discussions Miguel Toro, Daniel Gianola, Gustavo de los Campos, Andrés Legarra throughout years.