GPFN

Genomic prediction estimates the breeding value of an individual from its genotype alone, so that breeders can select parents before the plants or animals are ever phenotyped. The field has been dominated for two decades by linear mixed models such as GBLUP and its Bayesian relatives; deep neural networks, despite many attempts, have generally matched rather than beaten them. GPFN — the Genomic Prior-Data Fitted Network, from the Global Institute for Food Security — changes the setup rather than the architecture, and in doing so posts consistent gains over the linear baseline.

A Prior-Data Fitted Network is trained once, on data drawn from a generative prior, to approximate the posterior predictive distribution of that prior. GPFN's prior is a forward simulator of breeding populations: hundreds of thousands to millions of synthetic plant and animal populations are generated, each with its own founders, mating scheme, marker panel, trait heritability, and number of quantitative trait loci. The network learns to map a set of genotyped-and-phenotyped individuals plus a set of unphenotyped candidates directly onto breeding values for the candidates. Because that mapping is amortized across the whole prior, a trained GPFN is applied to a real population in a single forward pass, with no fitting, no hyperparameter search, and no per-population retraining. It is the genomic-prediction counterpart of the synthetic-prior approach used by TabPFN-Wide for wide omics tables and PerturbPFN for single-cell perturbation response.

The work was published in IEEE Transactions on Computational Biology and Bioinformatics in 2025, following a bioRxiv preprint that was revised through May 2025.

#Key Features

• No training at inference time: A breeder supplies a genotyped training set with phenotypes and a set of selection candidates; the model returns genomic estimated breeding values in one inference pass, avoiding the model fitting that conventional genomic prediction requires for every new population.
• Purely synthetic pretraining: Nothing in the pretraining corpus is real data. Populations are simulated with randomized heritability, QTL counts, dominance, founder sizes, and marker densities, which lets the prior cover regimes where real training data is scarce.
• Three released checkpoints: Pika targets nested association mapping and between-family prediction, Mongoose targets unstructured diverse populations, and Wombat implements a published soybean genomic-selection scheme. Each is a 311M-parameter model shipped in the same Docker image.
• Standard breeding file formats: VCF and HapMap genotypes plus CSV phenotypes are parsed into the model's binary input format by a bundled utility, so the pipeline drops into existing breeding workflows.

#Technical Details

GPFN is a transformer encoder that attends across individuals rather than across sequence positions: genotypes and phenotypes are separately embedded, the labelled training individuals are attended over as in-context examples, and an attention mask prevents query individuals from seeing one another's labels. Marker matrices are reduced by a dimensionality-reduction step before entering the encoder. The released checkpoints each carry 311 million parameters and roughly 1.16 GB of weights. Prior generation randomizes heritability between 0.2 and 0.8, QTL counts from 5 to 100, and marker panels from roughly 7,000 to 36,000 SNPs across populations of up to a thousand individuals evolved over tens of generations.

On three populations spanning two crop species and 16 traits, GPFN was significantly more accurate than the linear baseline on 13 traits. On a harder between-families structured prediction task in a third crop species, it matched the linear baseline overall and exceeded it at one location — a setting where classical models are known to be strong and where the authors do not claim an advantage.

#Applications

The immediate users are plant and animal breeding programs running genomic selection: cereal, oilseed, and legume breeders choosing parents from a nursery, and livestock programs ranking selection candidates. The zero-tuning deployment is most valuable for programs that lack the statistical-genetics staff to fit and validate a bespoke mixed model for each trait, and for early-cycle populations where the training set is small. The Docker distribution means a breeder needs only genotype and phenotype files, not a GPU-programming environment.

#Impact

GPFN demonstrates that the prior-data fitted network paradigm, first developed for small tabular machine-learning problems, transfers to quantitative genetics — a domain where the generative process is unusually well understood and can therefore be simulated faithfully. That is the key enabling condition, and it makes genomic prediction one of the cleanest test cases for simulation-based amortized inference in biology. The main constraints are that accuracy is bounded by how well the simulated prior matches the real breeding population, that the released checkpoints target specific population structures rather than being universal, and that reported gains come from a modest number of populations and traits. Code and weights are released under GPLv3.