← Latest papers
🧬 genetics

Is the Bitter Lesson coming to genomic prediction? An untuned tabular foundation model matches or beats specialized methods

This study demonstrates that the general-purpose, untuned tabular foundation model TabPFN-3.5 matches or outperforms specialized genomic prediction methods across multiple plant species and traits, supporting Sutton's "Bitter Lesson" that scalable computational approaches can surpass domain-specific human knowledge.

Original author: Flagel, L.

Published 2026-11-15
📖 5 min read🧠 Deep dive

Original author: Flagel, L.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Farmers and breeders have long faced a difficult choice: wait years to see how a plant or animal will perform in the real world, or try to guess its future based on its genetic code. In modern breeding, scientists can read the DNA of thousands of candidates quickly and cheaply, but measuring the actual traits—like how much milk a cow produces or how tall a corn stalk grows—remains slow and expensive. To bridge this gap, they use a tool called genomic prediction. This method trains a computer to learn the relationship between an organism's DNA and its physical traits using a group of known individuals, then applies that knowledge to predict the traits of new individuals who have only been tested for their DNA. For decades, the most successful tools for this job have been specialized models built by experts who carefully designed them to fit the specific rules of genetics, assuming, for instance, that traits are the sum of many small genetic effects.

A new study asks a provocative question: what if we stopped trying to outsmart the data with complex, custom-built rules? Instead of designing a model specifically for genetics, could a general-purpose computer program, trained on millions of synthetic examples of any kind of data, learn to predict genetic outcomes just by looking at the raw numbers? The researchers tested this idea using a powerful new type of software called a foundation model. Unlike traditional tools that require experts to tune every setting and make assumptions about how genes work, this model arrives pre-trained and ready to work. The team fed it the genetic data of plants and trees directly, without any special adjustments or genetic knowledge, to see if a "one-size-fits-all" approach could match or beat the specialized methods that have been refined over twenty years.

The researchers put this general-purpose model to the test against eighteen different prediction methods across six plant and tree species, including soybeans, rice, maize, and white spruce. They used a massive collection of data containing thousands of individuals and hundreds of thousands of genetic markers. In the first round of testing, they compared the new model's performance against the average results of the eighteen established methods reported in previous studies. The general-purpose model came out on top, ranking first in eleven out of the eighteen specific traits tested. It performed so well that it achieved the best overall rank across all the tests, suggesting that a tool designed for general data analysis could outperform tools built specifically for genetics.

To ensure the comparison was fair, the researchers then ran the most successful traditional models on the exact same data splits used by the new model. This step was crucial because previous studies had used different ways of dividing the data, which can sometimes make results look better or worse than they truly are. Even with this strict, head-to-head comparison, the general-purpose model held its ground. It significantly outperformed the best traditional models on four specific tests, including all three traits measured in soybeans and the flowering time of rice. In the soybean tests, where the dataset was the largest, the new model was clearly superior. However, it was not a perfect victory; on one test involving the wood density of white spruce, the traditional model remained more accurate.

The study also explored a critical limitation: how well these models work when predicting traits for families of plants that were not part of the original training group. In breeding, it is common to train a model on one set of families and then try to predict the performance of entirely new families. When the researchers tested this scenario with soybeans, the advantage of the new model disappeared. Both the general-purpose model and the traditional specialized models struggled equally when faced with these unseen families. This finding suggests that the new model's success relies heavily on recognizing patterns within the specific groups of relatives it has already seen, rather than discovering a universal rule that applies to all genetic lineages.

The researchers concluded that their results support a concept known as the "Bitter Lesson," which argues that in artificial intelligence, methods that simply scale up with more computing power eventually surpass those built on human intuition and domain-specific knowledge. The general-purpose model did not need to be told how genes work or how to handle genetic data. It simply processed the information it was given, leveraging its massive pre-training on synthetic data to find patterns that human-designed models had missed. While the improvements in accuracy were modest in absolute terms, they were consistent and significant in key areas. The study indicates that the future of genomic prediction may not lie in building more complex, specialized models, but in using powerful, general tools that learn directly from the data itself, leaving the heavy lifting to computation rather than human expertise.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →