Skip to content
Have you ever seen drone images of a breeding trial in an open field? I am always amazed by the beauty of these pictures. I love to see plants in such a perfect geometric disposition.
But let’s not get distracted… the whole story behind breeding trials is not about beauty. It is about science!
We should never forget that a breeding trial is a scientific experiment. Nothing more, nothing less. It is essentially something we establish to test a hypothesis, or a set of hypotheses.
So, the real question is: are breeding trials effective scientific experiments?
Let’s start from the basics. At university, I was taught that an effective experiment needs a robust experimental design. So, are breeding trials always planned with one in place?
Well… it depends. In academic research, yes! If you do not have a design, your paper will be rejected two minutes after you submit it to a journal. But in industry, there is no journal editor stopping you. I would guess that many traditional breeding programs worldwide use no formal experimental design at all.
Now, let’s take a step back. What do we actually need for a proper design?
I follow the famous “3 Rs” principles: Randomization, Replication, and Reduction of noise (also known as local control or blocking).
In this piece, I will focus on the first: randomization.
Randomization is statistically needed but operationally seen as chaos.
Randomization is the common thread across experimental designs in plant breeding. If you are implementing design in a program, the biggest mindset shift is not convincing people that design 1 is better than design 2. It is convincing them to move from non-randomized trials to randomized ones… and that’s a tough one! Let me explain why.
Dig into a trial dataset and you will quickly find out that breeding trials are often perfectly organized. Why? Because plant breeders love order. There is nothing we can do about that.
Suppose we have 100 plots in an unreplicated setting. So, there are 100 unique entries, either hybrids or lines. How are they assigned to a position in the field?
Very often, they are assigned following a specific order. For instance, they could be organized by “blood groups” (by which I mean by pedigree). It is common to see hybrids that share a pedigree planted in a row, one after the other. The same logic applies to line screening: sister lines sit close to each other.
Why is that?
As a breeder, there is always the need, and the (nice) curiosity, to learn and read the genetics by eye. If everything is well organized by blood group, you can walk into the field, look at groups of plots together, and infer whether the mother or the father of those crosses was good, or pick the best among sister lines. In this learning by order, your attention is not really on a single plot, but on a whole multi-plot picture.
That’s why randomization is the hardest thing to introduce when moving from traditional to data-driven breeding.
Why? Because you are basically messing up an order that a plant breeder has built and followed for many years. In the name of stats, you are asking someone who has already marketed many winning varieties to introduce chaos into the realm of order.
Randomization is about (a little) peace of mind.
Plant breeders love order, but the truth is that genetic differences thrive in chaos!
What do I mean?
It can be dangerous to place genetics in pedigree order without randomizing them in the field.
Let’s suppose we are selecting for a quantitative trait with low heritability in an unreplicated open field. If there is a field gradient and a group of full sibs (order) falls in a bad spot, we may discard them all, convinced it is bad genetics. But if we randomize instead, we increase the chances of sampling the field gradient. In fact, with randomization everyone gets a chance to land in a different spot, and siblings are split apart across the field. This reduces the risk that genetic differences become confounded with field variation.
Now imagine that lines from genetic background A are planted together in one part of the field and lines from genetic background B in another. Based on what we observe by eye, we discard several lines from genetic background A and advance several lines from genetic background B. Suppose that some of the discarded A lines would have become successful commercial varieties. Because they were planted in an unfavourable part of the field, they appeared inferior and were wrongly discarded.
Discarding those lines from genetics A statistically means we failed to detect a true positive. Economically, it means real money lost. Do we really want that?
So, the message is: if we do not randomize, we have zero peace of mind! If something goes wrong (a field gradient), at least randomizing means we did something to protect against it!
Randomization needs a data-analysis pipeline.
Why does the resistance to randomization run so deep? In my opinion, the key to getting randomization adopted is separating two things in time and space: evaluation and selection.
Let me be clear about what I mean.
Evaluation is about scoring plants, the phenotyping. It can be done by eye, or by a machine in more advanced setups. Selection is about picking the winners and advancing them to the next stage of the program.
In traditional programs, evaluation and selection happen simultaneously, through the breeder’s eyes. That’s exactly why order makes operational sense: similar genetics sit together, you look at them as a group, and while scoring (evaluation) you’re already choosing winners (selection). So the order is a tool that makes simultaneous eye-selection possible.
But should we really select this way?!
We need a change. We need to separate evaluation from selection.
Once field evaluation is completed and selection decisions are supported later by analysis, everything changes! You walk into the field and focus on a single plot, score it, and move on. You don’t lean over to compare it with a related plot next door, because you’ll see those differences later. Order in the field stops being necessary, because the comparison no longer happens through your eyes.
But if we separate evaluation and selection in time and space, something needs to sit in the middle. Right?
Yes, data analysis!
Why a data pipeline is the real precondition for randomization
You can only decouple evaluation from selection if something supports the selection process later: a data analysis pipeline!
If there is no analysis running (no software, no easy-to-use pipeline where breeders can upload their files and get results back), then randomization is nearly impossible to adopt. And the reason is simple: without that pipeline, there is no “later.” Selection will likely happen in the field, in real time, by eye. And to select by eye, you need order.
So, the resistance to randomization is not stubbornness. It is a missing tool, and sometimes missing knowledge. Not every breeder has had specialist training in statistical modelling, nor should every breeder be expected to build an analytical pipeline alone. Take that away, and field order becomes a rational response to a missing piece of infrastructure.
This is why it is now strategic for any private breeding business to have strong data analysis and quantitative genetics expertise on staff, or reliable consultants to call on.
Randomization needs adaptation.
From the throne of stats, it is easy to declare that we must randomize. Operationally, it takes an adaptation phase, and it is worth being honest about that.
Introducing randomization can cause friction across the team. The breeder must adapt to the chaos. The people supporting the program, those preparing sowing bags and organizing the layout must adapt too. You may run into headaches with mixed maturity groups that before were ordered, harvesting logistics, and a dozen other practical things…
It’s not an easy switch. But once you start, I am sure you see the benefit!
The best feedback I have heard is from breeders who have made the change (randomization). One told me he feels far more efficient. He now focuses on one plot at a time, scores it, and moves on. No longer losing focus by glancing at neighbouring plots, no longer hunting for patterns by blood type, and no longer swayed by his favourite pedigree!
The breeder can then wait for the statistical analysis to identify the strongest candidates!
Randomization is about risk management.
Statisticians often explain randomization in terms of valid statistical inference. The crude reality is that breeders sometimes feel detached from statistical inference. I always try to sell the story by saying that randomization is not only about stats, but it is also about money! This is a more practical way to look at it. Randomization protects breeders from making costly decisions based on field variation rather than genuine genetic differences (risk reduction). Isn’t that about money?
Let me close.
After all this, I have to disappoint you slightly. Randomization is just one piece of the puzzle. A very good one, but not enough on its own.
Remember the 3 Rs? We have only covered the first. Replication and the reduction of noise are still waiting.
But that is a story for another day.
If you’d like to discuss data analytics tools for your pipeline, please get in touch with the VSNi team. Our team is happy to run through the suite of tools and consultancy support on offer.
