Genomic prediction pricing has turned into a rumor market. The figures you hear at breeding conferences, and will hear in Valencia at Euroseeds, span two orders of magnitude — and both extremes can be true. Neither helps you, because genomic prediction isn’t one product.
What you’re actually buying
Vendors sell different things under the same name, from tools that answer within the hour to foundation models trained on pooled data and pitched for any crop. Computomics fits a separate model to each breeding program. Each model is trained on that program’s own genotype, phenotype, and environment records and retrained every season.
Each approach has its uses, and a foundation model may give a data-poor program a starting point. But ask what the model making your selection decisions learned from, and if it’s pooled data, whose data it was. A model fitted to your program learns your germplasm in your environments and belongs to you. That takes longer than an hour to set up.
This also explains the rumor market. A model built for U.S. hybrid corn is worth nothing to a tomato breeder in Almería, and its price is no guide to theirs.
Price it against your own trialing budget
A plot in a mechanized cereal trial and a hand-harvested vegetable plot with lab phenotyping cost wildly different amounts. So work in ratios. Add up a year of trialing, including land, labor, seed, phenotyping, and the season itself. A prediction layer pays for itself by removing entries that were never going to advance, so estimate what share of your entries a model would let you drop before planting. If the model costs less than that share of your budget, it’s worth having.
Culling is the easiest lever to calculate, but it’s rarely the only one. Add the value of reaching the market a season or two earlier. Then add what you’d otherwise spend building prediction infrastructure and staffing it.
If none of these add up, the program usually wasn’t ready for a model.
The readiness question
This part of the conversation gets skipped, and it’s why many prediction projects fall short.
Genomic prediction needs structured records from more than one season. Every observation has to link to a plot, a genotype, and the environment, treatment, or management it was grown under. One spreadsheet per season won’t support a model, and neither will 1-to-9 scores that can’t be traced to a plot.
Here’s a rough check: could you pull a few seasons of linked trial results in an afternoon? If it would take a month of cleanup, you know where to start. Without that linkage, no algorithm will rescue you. Data foundation work is less exciting than an accuracy chart. It’s also what separates a model that works in your program from a demo that works only on a slide deck.
Once the linkage exists, two seasons can be enough. A method called active learning uses a running model to point to the entries and measurements that would add the most information. A standard season with mostly new lines adds one more year effect. It teaches the model little about how the same genotypes respond from one season to the next, and that’s the signal it needs most.
Where the accuracy comes from
Messy data is the most common reason prediction underdelivers. The algorithm is further down the list. In a 2019 benchmark in G3, a team led by Michigan State University tested 12 algorithms on 18 traits in six species. The lineup ranged from rrBLUP to neural networks. No algorithm won everywhere. The neural networks weren’t the best for a single trait, and a simple average of several consistently matched or came close to the best one. Benchmarks like that hold the data fixed on purpose. In that setting, the algorithm is the easiest part of the stack to swap.
The gains come from joining genotypes, trial results, weather, soil, and other -omics layers. In published multi-environment studies, adding environmental data regularly improves predictions and lets a model predict environments it hasn’t seen. Linear models can use it too, as Diego Jarquín and colleagues showed with Arvalis in 2014. The catch is that someone must choose which interactions to model in advance, and that gets unwieldy as data layers pile up. Working out which stretch of the season matters for which trait is the part worth paying for.
Environments a model hasn’t seen include future ones. Since 2020, Computomics has predicted variety performance for any plot in the world. It can also predict across simulated climates 15 to 30 years out, including the many ways a season could go by then. That’s when varieties from today’s crosses will be on the market.
Keep your breeding system
A program that’s ready already has a breeding management system for its pedigrees, crosses, field designs, selections, and inventory. A prediction layer has no business replacing it.
What works is a layer on top. It pulls data together in the shape a model needs and sends predictions back into the tools your breeders already use.
Six questions to put to any vendor:
- Can you connect to my breeding system? Make sure the connection works both ways, pulling data in and sending predictions back.
- Where does the data live? For European breeders, germplasm data is the asset. Your own cloud account, or EU hosting where you own the data and models, is a very different deal from a platform that learns from your program.
- Do you reuse client data across clients? Without your explicit agreement, your germplasm may be training models your competitors pay for.
- Can I delete it? You want everything purged on request, with proof.
- What do I keep if I leave? Ask about your data, the fitted models, the cleaned datasets, and the prediction history. If you keep none of it, leaving means paying twice for cleanup.
- Will you still be around in five years? Models are retrained every season, so ask about funding and runway.
“Your data and your models belong to you, and we never use them for another client,” Schultheiss says. When breeders choose to pool their data, which can improve accuracy, Computomics can run the pool. Because the company doesn’t breed or sell seed, it has no varieties of its own to favor.
What a first engagement looks like
Start with one program, one crop, and a question with a clear answer. Fit a model on historical data and check its predictions against results you already know. Let that accuracy decide whether the work scales.
Once it scales, measure the payoff in time. One global brewer’s ESG reporting describes using Computomics’ predictions alongside traditional barley breeding. Early indications suggested the rate of improvement in key traits, including yield, could rise by 50 percent or more.
The first project also shows what working with a vendor is like. That often means a scatter of emails, shared drives, and procurement channels. Computomics is designing a single customer platform so you can see where every project stands without having to ask.
“Growth shouldn’t turn a working relationship into a ticket queue,” Schultheiss says. “We’re building it so the process runs the same way every time and you still know who you’re working with.”
Computomics has completed more than 200 projects in 18 countries since 2012 and processes client data in its own data center in Germany. It closed a €6.3 million Series B in August 2026. One pattern has held across its work: the programs that get value fixed their data first and kept their breeding system.
Computomics is based in Tübingen, Germany, and builds genomic prediction and breeding data infrastructure for commercial plant breeders. Find the team at trade table 134 at the Euroseeds Congress in Valencia, October 25 to 28, or book a consultation ahead. For our solutions, see computomics.com.
