AI-Driven Variant Discovery & Genomic Prediction

A genome contains millions of variants, but only a small fraction actually change a trait. Breeding and gene editing both depend on finding those impactful variants, and that discovery is the rate-limiting step. There are three broad strategies, each with a different trade-off:

Association testing (genome-wide association studies) is powerful but blunt: because nearby variants are inherited together, a significant signal points to a recombination bin that can span hundreds of thousands of DNA bases (often ~300 kb), not the causal base itself. We can see where an effect is, but not which variant causes it. That missing resolution is exactly what computational, AI-based screens promise to recover, cheaply.

Animal breeding predicts an animal's genetic merit (its breeding value) using the mixed model y = Xβ + Zu + e: observed performance y is split into fixed effects β (herd, year, season), random genetic merit u (the breeding values we want), and residual e. Genomic prediction replaces or augments pedigree with thousands to millions of DNA markers, building a genomic relationship between animals.

Marker data are “wide”: far more markers (p) than animals (n), the “big p, small n” problem. Ordinary regression fails, so we use regularised methods (genomic BLUP, Bayesian models) that shrink marker effects. Standard genomic prediction implicitly assumes every marker contributes a little. But biology says otherwise: a handful of variants matter a lot, most matter not at all. If we could tell the model which variants are likely impactful, predictions should improve, and that is where evolutionary scores come in.

Evolution has already run a billion-year experiment. Positions in DNA or protein that are essential are kept the same across species by negative (purifying) selection, mutations there are harmful and get removed. Positions that don't matter drift freely. So conservation is a clue to function.

We quantify this with a multiple-sequence alignment (MSA): stack the same gene from many species and look down each column. A change at a highly conserved column gets a high evolutionary score, it is “abnormal” given evolution and therefore likely impactful. Across crops, variants flagged this way are enriched for real fitness effects (shown in sorghum and potato biomass studies).

The same large-language-model technology behind modern AI can read biological sequence. Start with language: a model trained on text learns which word is expected next. In “a small caterpillar is born hungry and ____ everything in sight,” the model assigns eats 70%, tastes 25%, reads 4%. A change from “eats” to “reads” is flagged as abnormal, semantically wrong.

Protein language models do exactly this for amino-acid sequence. Trained on millions of natural protein sequences, they learn which residue belongs at each position. At a given site the model might predict R 80%, Q 10%, H 8%, L 2%. A mutation R→L lands on a very low-probability residue, so it scores as abnormal and likely impactful, without any alignment or labelled training data.

Do these scores actually help? Yes. When variants are weighted by their evolutionary scores in genomic prediction, accuracy for fitness-related traits improves across species: up to ~20% for grain yield in maize, ~10% for cassava biomass, and up to ~25% for biomass in potato. Telling the model which variants to trust beats treating all markers equally.

The bigger prize is new genomic techniques (NGT) such as precise single-base editing. To edit, you must know the exact causal base, the resolution GWAS lacks. Evolutionary scores can prioritise candidate single-base edits in silico, which are then checked by experimental mutagenesis. Proof-of-concept work in the model grass Brachypodium shows that mutations with the highest evolutionary scores have the largest effects on traits such as seed weight, evidence that AI can point editors at the right base.

Key takeaways

• Finding impactful variants, not just associated regions, is the real bottleneck in genomic breeding.
• Evolution conserves what matters; conservation and language-model scores turn that into a usable, genome-wide signal.
• Weighting variants by evolutionary scores measurably improves genomic prediction.
• The same scores may one day guide precise single-base edits for next-generation breeding.

Discussion questions

Glossary

Further reading & resources

• Ramstein, G. & Buckler, E. (2022). Prediction of evolutionary constraint... Genome Biology.
• Long et al. (2023). Frontiers in Plant Science, evolutionary-score-weighted prediction in cassava.
• Wu et al. (2023). Cell, deleterious variants and prediction in potato.
• Gianola, D. & Rosa, G. J. M. (2015). One hundred years of statistical developments in animal breeding. Annual Review of Animal Biosciences 3:19–56.
• Free courses at the FAO e-learning Academy.