DeepFARM

Most trait-associated variants in farmed animals fall outside coding sequence, which means genome-wide association studies routinely implicate regions whose function is unknown. In human genetics this gap has been narrowed by sequence-to-function models — Basenji, Basenji 2, Enformer — that learn to predict chromatin and expression signal from DNA sequence and can then be queried with a mutated sequence to estimate a variant's regulatory effect. Livestock and aquaculture genomes had no such models, largely because the functional genomics data to train them did not exist until the FAANG and AQUA-FAANG consortia produced it.

DeepFARM, from the Centre for Integrative Genetics at the Norwegian University of Life Sciences with AquaGen, fills that gap for four species: cattle, chicken, pig, and Atlantic salmon. Each species gets its own network trained on its own chromatin profiles, and the trained networks are then applied without retraining to score every non-coding variant of interest genome-wide.

The framework is deliberately conservative in its modelling choices. It benchmarks two established architectures — DeepSEA and DanQ — rather than proposing a new one, and reports that DanQ, a convolutional layer followed by a bidirectional LSTM, consistently wins.

#Key Features

• Per-species models: Four independent networks rather than one multi-species model, each trained against that species' assembly and chromatin data, so predictions reflect species-specific regulatory grammar.
• Multi-assay chromatin targets: Training labels span ATAC-seq, DNase hypersensitivity, CTCF ChIP-seq, and histone modification ChIP-seq, giving each 200 bp bin a multi-label output vector.
• In silico saturation mutagenesis: Substituting every base in a window and re-scoring localises which positions drive a predicted regulatory change, which the authors use to interpret candidate causal variants.
• Motif recovery: Of the 320 convolutional filters learned per model, 49, 51, 63, and 34 match known transcription factor motifs in cattle, chicken, pig, and salmon respectively, evidence that the models learn real regulatory syntax rather than assembly artefacts.
• Nextflow pipeline: Training, tuning, and inference run as a containerised Nextflow workflow, so the whole framework can be retrained on a new species given peak calls and a reference genome.

#Technical Details

Chromatin peaks come from FAANG for the terrestrial species and AQUA-FAANG for salmon: 95 profiles for cattle, 80 for pig, and 97 for chicken across liver, lung, spleen, skeletal muscle, subcutaneous adipose, cerebellum, brain cortex, and hypothalamus, and 301 profiles for Atlantic salmon across brain, liver, gill, gonad, and muscle. Genomes (ARS-UCD1.2, Sscrofa11.1, GalGal6, Ssal v3.1) are divided into 200 bp bins; each bin is one-hot encoded together with two flanking bins on each side to form a 1,000 bp input, and labelled positive for a profile when at least half the bin overlaps a peak. One chromosome is held out for validation and another for testing.

The best DanQ models reach median AUROC of 0.9110 in cattle, 0.9065 in salmon, 0.8825 in chicken, and 0.8512 in pig, in each case a little above the corresponding DeepSEA model. Downstream evaluation is more mixed than the abstract's summary implies. Classifying fine-mapped eQTL causal variants across 13 pig tissues with a random forest over the functional scores gives a mean AUROC of 0.5726, ranging from 0.7469 in spleen down to 0.5226 in uterus. In the Atlantic salmon genomic prediction case study, a 9,073-SNP subset chosen by functional score matches the full 70k array (predictive correlation 0.404 versus 0.421 under a Bayesian method, 0.365 versus 0.373 under GCTA) while clearly beating random subsets of the same size (0.359 and 0.348).

#Applications

The immediate use is prioritising candidate causal variants under a GWAS peak in a breeding programme, where the difference between a tagged variant and the functional one determines whether a marker transfers across populations. The salmon result suggests a second use: assembling a small, functionally enriched marker panel that performs like a much denser array, which matters when genotyping cost scales with panel size across hundreds of thousands of animals. The saturation mutagenesis output supports mechanistic follow-up, as in the cattle analysis of a variant colocalised with the DGAT1 cis-eQTL and a protein yield association. The Nextflow pipeline makes the same treatment available for any species with chromatin peaks and an assembly.

#Impact

DeepFARM brings the sequence-to-function paradigm into livestock and aquaculture genomics, where the reference and functional annotation resources have lagged human genetics by a decade. Its honest reporting of modest eQTL classification performance is a useful calibration for a field where such models are often assumed to transfer cleanly. Pretrained DanQ weights for all four species are published as downloadable .h5 release assets and are explicitly intended for regulatory impact prediction without additional training; the code repository, however, ships no LICENSE file, so reuse terms are unstated. The work appeared as a preprint in February 2025 and was published in NAR Genomics and Bioinformatics in October 2025.