Work

LifeDive: Cross-kingdom phenotype estimates for 35,856 species

Year: 2026
Type: preprint
Abstract: Dense, standardized species-by-trait matrices support comparative biology and ecology, but theircoverage remains uneven. LifeDive uses a general-purpose, web-search-enabled language modelwithin a scripted annotation pipeline to estimate traits across animals, plants, and fungi. Theworkflow assembled 35,856 species by 196 phenotype questions, yielding 7,027,776 entries in acompleted matrix; 16.6% of entries were imputed from other machine-derived annotations. Fourordered substantive responses express model judgments about trait statements, with a separatecode for unknown or inapplicable cases. An exploratory 14-factor representation summarized broadbiological contrasts, although statistical factor-retention criteria suggested substantially moredimensions. A benchmark screened 168 external criteria, with selected results reported for 44;many biologically direct associations were strong, particularly for taxonomy and selected ecologicaltraits. A separate model-assisted audit reported 96.5% directional agreement in a selectedtaxonomy-rule subset and approximately 237 agreements among 240 selected species-specificchecks. These audit subsets do not estimate accuracy across the whole matrix, and theAmphiBIO-derived benchmark results require qualification because unrecorded categories weretreated as negative indicators. Trait profiles also produced recognizable taxonomic and functionalgroupings. LifeDive offers a broad computational phenotype resource for exploration, hypothesisgeneration, and prioritizing expert review. Individual values remain estimates, and downstreaminference requires attention to imputation, sampling, source dependence, and trait-specificverification. The reported application programming interface and computing expenditure wasapproximately US$10,000.
Language: English
Open Access status: green
OpenAlex ID: W7225864973
Cites:
Cited by:
Related to: 10