Preprint · not peer-reviewed Researchers Educators Students

Preprint: untuned tabular foundation model matches or beats specialist methods across 18 genomic prediction tasks

A preprint benchmarks TabPFN-3.5, a general-purpose pretrained model with no genetics-specific design, against 18 published genomic prediction methods across 18 traits in six plant species — finding it ranks first on eleven traits with no tuning.

Published · AI-drafted summary based on 1 public source
Illustration for generic story
Illustrative image — not from the source article.
Share

A preprint posted to bioRxiv asks whether a general-purpose tabular foundation model — TabPFN-3.5, trained without any genetics-specific assumptions — can compete with specialist genomic prediction methods when given only marker genotype data. The authors benchmark the model against a published comparison of 18 genomic prediction approaches across 18 quantitative traits in six plant species, using the same evaluation structure where possible, and report that TabPFN-3.5 ranked first on 11 of 18 traits and achieved the best median rank across all comparisons.

The authors frame this as a contribution to a broader question in machine learning — sometimes called the "bitter lesson" — about whether domain-specific inductive biases continue to confer an advantage as general-purpose models scale. In genomic prediction, methods such as GBLUP and BayesB incorporate explicit assumptions about the genetic architecture of traits; that an untuned foundation model can match or exceed these on plant breeding benchmarks is a methodologically significant observation.

The authors note that the published benchmark they used did not specify its test sets, so they re-ran GBLUP and RR-BLUP as internal controls to validate comparability. The preprint has not been peer-reviewed. Researchers in quantitative genetics, plant breeding, statistical genomics, and those following the intersection of large language and foundation models with biological data will find the benchmarking analysis of direct interest.

Sources

Read the original reporting — these are the public sources this summary draws from.

  1. Primary sourcePreprint bioRxiv (Cold Spring Harbor Laboratory) · 2026-11-15
    Is the Bitter Lesson coming to genomic prediction? An untuned tabular foundation model matches or beats specialized methods

Tags

genomic-prediction foundation-model tabpfn machine-learning quantitative-genetics plant-genetics gblup preprint
Share

About Genetic Current

Educational summaries of public genetics news

Genetic Current is the news section of Evagene, an academic, research, and educational pedigree-modelling platform. Stories are AI-drafted summaries of items from trusted public sources, written for researchers, clinicians, educators, students, genealogists, and patients with an interest in genetics. Summaries are for educational and research purposes only and are not medical advice.

Join the Evagene Alpha Waiting List