@druncie.bsky.social
These are great!
January 14, 2026 at 4:37 PM
These are great!
What do you think?
- Are there other ways the Genomic Prediction models can be used?
- Has the field just "given up" on Recurrent Genomic Selection and decided that focusing on intensity is good enough?
- Am I missing Genomic Selection success stories in plant breeding?
12/n
- Are there other ways the Genomic Prediction models can be used?
- Has the field just "given up" on Recurrent Genomic Selection and decided that focusing on intensity is good enough?
- Am I missing Genomic Selection success stories in plant breeding?
12/n
January 9, 2026 at 11:58 AM
What do you think?
- Are there other ways the Genomic Prediction models can be used?
- Has the field just "given up" on Recurrent Genomic Selection and decided that focusing on intensity is good enough?
- Am I missing Genomic Selection success stories in plant breeding?
12/n
- Are there other ways the Genomic Prediction models can be used?
- Has the field just "given up" on Recurrent Genomic Selection and decided that focusing on intensity is good enough?
- Am I missing Genomic Selection success stories in plant breeding?
12/n
In the remainder of the paper, I study why accuracy falls apart more in some GS schemes than others and provide mathematical and graphical diagnostic tools to predict in advance which types of breeding programs are more likely to be able to implement effective GS.
11/n
11/n
January 9, 2026 at 11:58 AM
In the remainder of the paper, I study why accuracy falls apart more in some GS schemes than others and provide mathematical and graphical diagnostic tools to predict in advance which types of breeding programs are more likely to be able to implement effective GS.
11/n
11/n
Based on these results, I conclude that cross-validation is not a useful tool for evaluating Genomic Prediction models.
We should be evaluating models based on how much they can improve genetic gain, not whether they are "accurate" in contexts that matter little to breeders.
10/n
We should be evaluating models based on how much they can improve genetic gain, not whether they are "accurate" in contexts that matter little to breeders.
10/n
January 9, 2026 at 11:58 AM
Based on these results, I conclude that cross-validation is not a useful tool for evaluating Genomic Prediction models.
We should be evaluating models based on how much they can improve genetic gain, not whether they are "accurate" in contexts that matter little to breeders.
10/n
We should be evaluating models based on how much they can improve genetic gain, not whether they are "accurate" in contexts that matter little to breeders.
10/n
More importantly, in the latter case, the cross-validation estimates are basically uncorrelated with the actual accuracy in most cases. Cross-validation cannot distinguish cases where Genomic Selection will work from cases where it will not.
9/n
9/n
January 9, 2026 at 11:58 AM
More importantly, in the latter case, the cross-validation estimates are basically uncorrelated with the actual accuracy in most cases. Cross-validation cannot distinguish cases where Genomic Selection will work from cases where it will not.
9/n
9/n
Here are results for a "typical" case similar to examples from the literature: a model tested in a diverse breeding population of ~400 lines:
Cross-validation underestimates accuracy for schemes targeting intensity or accuracy, but usually overestimates accuracy for schemes targeting speed.
8/n
Cross-validation underestimates accuracy for schemes targeting intensity or accuracy, but usually overestimates accuracy for schemes targeting speed.
8/n
January 9, 2026 at 11:58 AM
Here are results for a "typical" case similar to examples from the literature: a model tested in a diverse breeding population of ~400 lines:
Cross-validation underestimates accuracy for schemes targeting intensity or accuracy, but usually overestimates accuracy for schemes targeting speed.
8/n
Cross-validation underestimates accuracy for schemes targeting intensity or accuracy, but usually overestimates accuracy for schemes targeting speed.
8/n
So how "bad" are cross-validation estimates? I ran simulations to answer this question:
Say you test your GP model by cross-validation and estimate that it has an accuracy of r=0.50. How well is that model likely to work if you use it to improve a) intensity, b) accuracy, or c) speed?
7/n
Say you test your GP model by cross-validation and estimate that it has an accuracy of r=0.50. How well is that model likely to work if you use it to improve a) intensity, b) accuracy, or c) speed?
7/n
January 9, 2026 at 11:58 AM
So how "bad" are cross-validation estimates? I ran simulations to answer this question:
Say you test your GP model by cross-validation and estimate that it has an accuracy of r=0.50. How well is that model likely to work if you use it to improve a) intensity, b) accuracy, or c) speed?
7/n
Say you test your GP model by cross-validation and estimate that it has an accuracy of r=0.50. How well is that model likely to work if you use it to improve a) intensity, b) accuracy, or c) speed?
7/n
The problem is, most papers (including most of mine!) test their Genomic Prediction models using cross-validation in a way that simulates using it to increase selection intensity. The graphs above show that this is probably the least impactful use of Genomic Prediction models in plant breeding.
6/n
6/n
January 9, 2026 at 11:58 AM
The problem is, most papers (including most of mine!) test their Genomic Prediction models using cross-validation in a way that simulates using it to increase selection intensity. The graphs above show that this is probably the least impactful use of Genomic Prediction models in plant breeding.
6/n
6/n
However, only targeting speed can have a big effect on genetic gain. In moderate-sized breeding populations, the potential gains for targeting intensity or accuracy < 20-40% unless h2 is very low. But by targeting speed, the rate of gain could increase many-fold!
5/n
5/n
January 9, 2026 at 11:58 AM
However, only targeting speed can have a big effect on genetic gain. In moderate-sized breeding populations, the potential gains for targeting intensity or accuracy < 20-40% unless h2 is very low. But by targeting speed, the rate of gain could increase many-fold!
5/n
5/n
Based on the breeder's equation, Genomic Selection can improve genetic gain by increasing selection intensity (i), selection accuracy (ρ), or reduce cycle lengths (L).
But each of these requires applying Genomic Predictions to entirely different sets of breeding candidates:
4/n
But each of these requires applying Genomic Predictions to entirely different sets of breeding candidates:
4/n
January 9, 2026 at 11:58 AM
Based on the breeder's equation, Genomic Selection can improve genetic gain by increasing selection intensity (i), selection accuracy (ρ), or reduce cycle lengths (L).
But each of these requires applying Genomic Predictions to entirely different sets of breeding candidates:
4/n
But each of these requires applying Genomic Predictions to entirely different sets of breeding candidates:
4/n
But is showing that a Genomic Prediction model is accurate the same as showing that it is useful? Does this prove that it will make Genomic Selection work?
It really depends on how Genomic Selection is used.
3/n
It really depends on how Genomic Selection is used.
3/n
January 9, 2026 at 11:58 AM
But is showing that a Genomic Prediction model is accurate the same as showing that it is useful? Does this prove that it will make Genomic Selection work?
It really depends on how Genomic Selection is used.
3/n
It really depends on how Genomic Selection is used.
3/n
The literature is full of examples of highly accurate Genomic Prediction models developed for a wide range of crops.
Here are accuracy reports from a random sample of papers. These numbers look really good!
2/n
Here are accuracy reports from a random sample of papers. These numbers look really good!
2/n
January 9, 2026 at 11:58 AM
The literature is full of examples of highly accurate Genomic Prediction models developed for a wide range of crops.
Here are accuracy reports from a random sample of papers. These numbers look really good!
2/n
Here are accuracy reports from a random sample of papers. These numbers look really good!
2/n