bioRxiv · 10.64898/2026.11.03.764067
Is the Bitter Lesson coming to genomic prediction? An untuned tabular foundation model matches or beats specialized methods
Abstract
Genomic prediction (GP) is dominated by models designed around assumptions about how genetic variation affects phenotypes. Here we asked how well a general-purpose pretrained tabular foundation model called TabPFN-3.5 does with no genetics-specific modeling and no tuning when given marker genotypes directly. We benchmarked TabPFN-3.5 using a published comparison of 18 GP methods on 18 traits in six plant species. Against the published means, TabPFN-3.5 ranked first on 11 of 18 traits and had the best median rank. Because the published benchmark did not identify its test sets, we reran GBLUP, RKHS and elastic-net models on the same splits as TabPFN-3.5. Against the best of these, TabPFN-3.5 was significantly behind on only one trait, spruce wood density, and ahead on 12 of 18, with significant gains on all three soybean traits and rice flowering time. In the soybean panel, which had the largest number of lines, its advantage was clearest. In leave-families-out cross-validation, where models were trained on some soybean families and used to predict lines from other families withheld from training, every model lost accuracy and TabPFN-3.5 performed on par with GBLUP and RKHS. That a general-purpose learner matches or exceeds decades of specialized GP models is consistent with Sutton's Bitter Lesson, which holds that methods that scale with computation eventually overtake those built on human knowledge of the domain.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Flagel, L.. 2026-11-15. Is the Bitter Lesson coming to genomic prediction? An untuned tabular foundation model matches or beats specialized methods. https://doi.org/10.64898/2026.11.03.764067
Cite the original work for its findings. Save a collection to share your selection of sources.