Genomic selection & AI-assisted plant and animal breeding

A breeding cycle takes years because the true value of a line or animal is only visible after field trials or progeny testing. Genomic selection predicts breeding values from DNA markers, and AI adds environment modelling and image-based phenotyping on top. Breeders select earlier, test fewer candidates and shorten cycles; the work stays with the breeding team, not an off-the-shelf chatbot.

Problem & value

• Today: plant breeders run multi-year, multi-location trials and phenotype by hand or by eye. Livestock breeders rely on pedigree and progeny testing. Both are slow and costly, and climate change makes target environments move.
• AI value: since Meuwissen et al. (2001) genomic prediction estimates breeding values from genome-wide markers without waiting for phenotypes [1]. Modern pipelines add genotype-by-environment (GxE) modelling, drone and greenhouse imaging and optimised crossing. The main gains are shorter generation intervals and higher selection accuracy for hard-to-measure traits. Dairy cattle breeding is the best-documented example; sourced figures vary by species and programme, so benchmark on your own data rather than quoting averages.
• EU context: the NGT Regulation (EU) 2026/1388 entered into force on 16 July 2026 and applies from 17 July 2028 [2][3]. It creates a lighter route for "Category 1" gene-edited plants, which raises demand for fast, accurate selection and trait-prediction tooling.

Variants

• Plant breeding, field crops (wheat, maize, sugar beet, rapeseed): GxE prediction across trial networks, drone phenotyping, speed breeding.
• Horticulture and vegetables: greenhouse image phenotyping, many small markets, shorter cycles.
• Livestock (dairy, pigs, poultry, aquaculture): single-step GBLUP, herdbook integration, sensor traits (see PRED-044, VIS-014).
• Public and international programmes: CGIAR centres, national institutes and publicly funded pre-breeding, with open data and tools.
• Gene-edited material: AI prioritises edit targets and predicts effects (links to SCI-021, SCI-017).
• Small breeders and hobby/private: limited to marker-service offerings and simple software; full pipelines are not realistic.

How it works

1. genotyping (array / low-pass sequencing / imputation) + phenotypes (trials, sensors, drone & greenhouse images
2. CV models) + environment covariates (weather, soil, management)
3. data QC, pedigree/marker curation, imputation
4. prediction model (GBLUP / ssGBLUP / Bayesian / ML / deep, GxE terms)
5. GEBV + accuracy estimates per candidate
6. selection index & optimal-contribution / crossing design
7. breeder review
8. crosses, advancement, trial design
9. new phenotypes feed back to retrain model

• Prediction: linear mixed models (GBLUP, ssGBLUP for livestock with pedigree) remain strong baselines; gradient boosting, multi-trait and deep models help mainly with non-additive effects and GxE.
• Image phenotyping: segmentation and detection models turn drone or greenhouse images into plot-level traits; VLMs help with scoring and annotation.
• Optimisation: crossing and mating plans balance gain against inbreeding using solvers.
• Human-in-the-loop: breeders decide on selection and crosses; models give ranked proposals with accuracies. Prediction must be validated on independent years and sites.

Data preparation

Genotypes come from array or sequencing providers (VCF, PLINK files), trial data from field books, spreadsheets, trial-management software or paper, images from drones and phenotyping platforms, and pedigrees from herdbooks or breeding databases. Typical gaps: mismatched line or animal IDs between genotype and phenotype files, missing pedigree links, inconsistent trait definitions across sites and years, uncalibrated sensors, missing GPS or plot layouts, genotyping errors and missing markers, and unclear access rights between partners.

Preparation steps: build an ID registry and clean the pedigree (PREP-004, PREP-011); digitise paper field books and capture phenotypes digitally at source (PREP-006, PREP-008); define trait ontologies (PREP-012, PREP-019); run marker QC (call rate, MAF, imputation) with PLINK or bcftools; link images to plots with metadata (PREP-010); export from breeding databases (PREP-005); check legal basis where farmer or owner data appear (PREP-025). Tools: BreedBase or BMS, PLINK, Nextflow. See DATA-003 and GUIDE-003.

Expect 50-60 percent of total effort in harmonising IDs, trial data and genotypes.

Implementation paths

Which path for whom

• Enterprise: C - breeding data and pedigrees are core intellectual property, open genomic-prediction tools (rrBLUP, BGLR, sommer, ASReml alternatives) are mature, and a shared in-house platform serves many programmes; A suits small breeders without quantitative genetics staff.
• Developer: C - open statistical genetics and image-analysis libraries are the standard base.
• Public: C - research institutes and public breeding programmes use open tools and national infrastructure, which fits sovereignty and data-sharing rules.

Path A — Buy (off-the-shelf)

• Breeding data platforms: Options include BreedBase (open source, hosted by cooperating institutions), Breeding Management System (BMS, Integrated Breeding Platform), plus commercial breeding LIMS (verify current vendors and EU hosting).
• Genotyping and evaluation services: commercial array and sequencing labs and national evaluation centres (e.g. Interbull for international dairy evaluations [4]); EU-based providers are preferred for data location.
• Choose A when: you are a small or mid-size breeder without a quantitative-genetics team; expect limited model customisation.

Path B — Build on cloud APIs

• Cloud storage and compute (EU regions) for pipelines; vision APIs or general VLMs (Claude, Gemini, Mistral) to score images and clean trial notes; LLMs to assist with trial metadata, literature and code, not to predict breeding values.
• Managed ML (Azure ML, Vertex AI, SageMaker) for training tabular and image models.
• Choose B when: you have data scientists but little infrastructure.

Path C — Self-host (open source)

• Genomic prediction: rrBLUP, sommer, BGLR (R), BLUPF90 family (livestock), GCTA, plus scikit-learn, XGBoost and PyTorch for ML and deep models.
• Genotype processing: PLINK, Beagle for imputation, DeepVariant for sequencing variant calls.
• Image phenotyping: PlantCV, Segment Anything-class models, YOLO-family detectors, open VLMs (Qwen-VL).
• Data management: BreedBase for trials and pedigree.
• Hardware: HW2 is enough for most plant programmes; large livestock evaluations are CPU- and memory-heavy, so a big-RAM server matters more than a GPU.

Path D — Custom training

Worth it for large programmes with thousands of genotyped and phenotyped lines per year. Typical work: multi-trait GxE deep models, crop-specific image models, genomic language models fine-tuned on proprietary sequence data. The data requirement is high; start with well-tuned mixed models and require ML models to beat them in cross-validation before adopting them.

State of the art (2026)

• GBLUP and its single-step variants remain the production backbone; deep and ML models give mixed, trait- and population-dependent gains. Reports of large gains often rely on weak validation (random cross-validation instead of leave-site/leave-year-out).
• Foundation models: DNA and plant genomic language models and protein models are used for variant prioritisation and edit design; their value for within-programme breeding value prediction is still unproven versus classic methods.
• Phenomics: drone multispectral and greenhouse robots with CV-based trait extraction are routine at large seed companies and research centres.
• Speed breeding and genomic selection are combined to cut cycles; AI-designed crossing schemes are in active use.
• Limits: accuracy drops across unrelated populations and new environments; small training sets; phenotyping noise; ancestry-specific bias; selection reduces genetic diversity if inbreeding is not managed.

Shared resources & combinations

• Tabular ML and statistics stack shared with PRED-013 (yield prediction) and PRED-044 (livestock sensor prediction).
• Vision stack shared with VIS-005 (plant disease) and VIS-014 (livestock monitoring); one drone pipeline serves several uses.
• Genomics: SCI-009 (human variant interpretation) shares variant-calling tools; SCI-017 (multi-omics target ID) and SCI-021 (sequence design) are natural neighbours for gene-editing programmes.
• Hardware: HW2 GPU plus a high-memory CPU server.

Effort, cost & KPIs

• Path A: weeks to months; per-sample genotyping and licence costs dominate.
• Path B/C: 6–12 months for a validated pipeline; the largest costs are genotyping, phenotyping and staff, not compute.
• Path D: multi-year programme with dedicated quantitative genetics and data science team.
• KPIs: prediction accuracy (correlation of GEBV with realised performance) in leave-year/site-out validation, genetic gain per year, generation interval, number of field plots saved, cost per selected candidate, inbreeding rate, share of image-scored traits vs manual.

People & change

Involve the head of breeding and the breeders who make selection decisions from the start; they must trust and be able to question GEBVs. Add field and phenotyping staff early since they generate the data, and IT for storage and compute. Management owns IP and licence rules (variety protection, partner data). Suppliers such as genotyping labs and partner programmes need data-sharing agreements; the data owners of herdbooks and trial networks must release data.

Data protection: mostly no personal data; owner, farmer or staff identifiers can appear in herdbooks and trial records, so involve the data protection officer (DPO) and sign processor agreements (Art. 28 GDPR) with cloud providers; a DPIA is normally not required.

Works council: no in the typical case, as the tools evaluate plants and animals, not staff. If field apps or image tools record individual worker performance, section 87(1) no. 6 BetrVG applies, and the council is informed in good time about AI-planned work processes (section 90(1) no. 3).

AI literacy: the AI Act Art. 4 requires measures that support staff AI literacy; train breeders on cross-validation, accuracy per environment and the limits of prediction for untested populations. Pilot on one crop or herd trait and run selection in parallel with the classical method. Concern: "models override breeders' experience" - breeders keep the final selection and mating decisions.

Guides: GUIDE-008, GUIDE-009, GUIDE-005.

Risks, regulation & compliance

• EU AI Act: minimal risk; no Annex III category applies and AI literacy (Art. 4) covers staff [5]. Not a prohibited practice.
• GDPR: mostly no personal data; exceptions are farmer, herd-owner and trial-staff data in herdbooks and field records, and any human samples used as controls.
• Plant sector: NGT Regulation (EU) 2026/1388 (Category 1 vs 2, no use in organic production, seed labelling, patent rules) [2][3]; plant variety rights (UPOV, CPVO) and seed-marketing law.
• Animal sector: Regulation (EU) 2016/1012 on zootechnical standards governs herdbooks and genomic evaluations [6]; animal welfare law applies to breeding goals.
• Other: access and benefit-sharing for genetic resources (Nagoya Protocol), data ownership with farmers and partners, biosecurity for sequence data.
• Mitigations: independent validation, diversity monitoring, documented models and data lineage, contracts on data use.

Getting started

1. Audit what you have: genotyped and phenotyped individuals, trial sites and years, pedigree quality.
2. Fit a GBLUP baseline with rrBLUP or BLUPF90 and measure leave-year-out accuracy.
3. Add environment covariates and test one ML or GxE model against the baseline.
4. Pilot drone or greenhouse image phenotyping on one trait where manual scoring is the bottleneck.
5. Integrate predictions into the selection meeting and track genetic gain per year.

Sources

1. Meuwissen, Hayes, Goddard (2001), Prediction of total genetic value using genome-wide dense marker maps, Genetics — https://doi.org/10.1093/genetics/157.4.1819
2. European Commission — New genomic techniques — https://food.ec.europa.eu/plants/new-genomic-techniques_en
3. EFSA — New genomic techniques — https://www.efsa.europa.eu/en/topics/new-genomic-techniques
4. Interbull Centre — https://interbull.org
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. Regulation (EU) 2016/1012 (zootechnical standards) — https://eur-lex.europa.eu/eli/reg/2016/1012/oj
7. BreedBase — https://breedbase.org
8. PlantCV — https://plantcv.org