Reader Comments

Post a new comment on this article

Why did you study high heritability traits ?

Posted by EduMan on 18 Nov 2008 at 03:01 GMT

Lee et al. propose a method for model selection and prediction of future phenotypes. I agree with the authors on the value of DNA scans to improve accuracy (and precocity !) of prediction. But their application to traits with unusually high heritabilities (from 0.55 to 0.99) deserves some comments:

A future phenotype can be represented by P=G+E, where g are future gentic effetcs and E a future residual. When heritability is very high, as in Lee et al., the future residual term E will not influence the future phenotype, and high prediction accuracies can be reached.
But, in real life, most complex traits have heritabilities around 0.3. Prediction accuracies of future phenotypes for mice traits with low heritabilities can be found in Legarra et al (Genetics, 180:1-8).

Of course, in animal breeding applications the goal is the prediction of a large number of phenotypes of future progeny of a sire (or a dam). At the limit, the prediction accuracy of the MEAN PHENOTYPE of an INFINITE NUMBER OF FUTURE PROGENY does not depend on the residual variance.
But, in applications aiming at the prediction of future INDIVIDUAL PHENOTYPES of complex traits, accuracies of prediction may be lower.

RE: Why did you study high heritability traits ?

unehong replied to EduMan on 20 Nov 2008 at 08:48 GMT

The authors appreciate the feedback on our article. In response to the points raised:

The heritabilities that we have chosen are not necessarily 'unusually high'. For example, in humans there are many traits with high heritabilities, including diseases (on liability scale) and height.

We agree that the upper limit of phenotype prediction from genetic data is determined by the heritability, and mention that in the paper (Table 2, page 4 last paragraph). We therefore also agree with the general comment that prediction of phenotypes is more difficult with lower heritable traits. What is relevant in this study is that a large proportion of the genetic variation could be predicted.

We note that the accuracy of predicting phenotypes will be low even though estimated genetic values may be highly accurate for traits with a low heritability. Predicting unobserved phenotypes of lower heritable traits may be improved if we could get more handle on estimating non-additive effects (incl epistasis) and possibly GxE or epigenetic terms.
In our study we were successful in capturing dominance variation but in our data fitting epistasis did not help. It would need more work to understand when that can be successfully fitted.

The authors