🧬 biology

Cross-kingdom phenotype annotations for 35,856 species generated with a web-search-enabled language model

This paper presents a comprehensive cross-kingdom phenotype dataset comprising over 7 million annotations for 35,856 animal, plant, and fungal species, generated through a web-search-enabled large language model pipeline to facilitate broad-scale comparative biology and conservation research.

Original authors: Tyler M. Moore

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tyler M. Moore

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Imagine you are trying to build a massive, universal "encyclopedia of life" that describes every single animal, plant, and fungus on Earth. You want to know things like: Does it have a shell? Does it live underwater? Is it poisonous? Does it eat plants?

Usually, filling out this encyclopedia is like trying to paint a mural by hand, one tiny brushstroke at a time. Scientists have to read thousands of old books and scientific papers to find the answer for each species. It's slow, expensive, and often leaves huge gaps because some animals are famous (like lions) while others are obscure (like a specific type of mold), and the books just don't exist for the obscure ones.

The "LifeDive" Project: A Digital Librarian with a Superpower

This paper introduces a new dataset called LifeDive. Instead of hiring thousands of human researchers to read every book, the author used a "super-smart" computer program (a Large Language Model) that has a special superpower: it can search the live internet.

Think of this computer not just as a robot that reads a library, but as a digital detective that can instantly look up facts about a specific bug in a remote forest or a rare mushroom in a cave, just like a human would use Google.

How They Did It

The Target List: The team started with a master list of about 36,000 species (animals, plants, and fungi). They made sure to pick a mix of common and rare creatures so the data wouldn't just be about "charismatic" animals like pandas.
The Questionnaire: They created a standardized test with 196 questions covering everything from biology (cell walls, venom) to behavior (nocturnal, social) and even how humans see them (is it beautiful? is it a pest?).
The Detective Work: The computer was asked to answer these 196 questions for every single species on the list. Because it could search the web, it could find answers for species that no single human expert has ever studied in depth.
The Scale: The result is a massive spreadsheet with 7 million answers.

The "Confidence Scale"

The computer didn't just say "Yes" or "No." It used a 5-point scale to show how sure it was:

Definitely No
Probably No
Unknown / Not Applicable (The computer admitted it didn't know)
Probably Yes
Definitely Yes

This is crucial. It means the data tells you not just what the trait is, but how confident the computer is about that answer.

Cleaning Up the Mess

Because the computer is smart but not perfect, it sometimes guessed or missed things. To fix this, the researchers used a statistical "fill-in-the-blanks" tool (called random forest imputation).

Analogy: Imagine a crossword puzzle where some squares are blank. If you know the word "FISH" goes in one row and "GILLS" in a column, you can logically guess the missing letter. The computer did this for the missing answers, using patterns it learned from the other 195 questions to fill in the gaps.

Did It Work? (The "Spot Check")

The authors didn't just trust the computer; they put it through a rigorous test:

The "Gold Standard" Check: They compared the computer's answers against existing, trusted databases (like FishBase or PanTHERIA) for species that were already known. The computer's answers matched the human-curated data almost perfectly (often over 95% agreement) for big categories like "Does it have a backbone?" or "Does it have a shell?"
The "Expert Audit": They took 200 random species and had a different AI (acting as a second opinion) check the work. The two AIs agreed on the direction of the answers about 95–97% of the time.
The "Grouping" Test: They looked at the data to see if it made biological sense. Did all the fish group together? Did all the mushrooms group together? Yes. The computer naturally grouped similar creatures together, proving it understood the "shape" of life, even if it wasn't a biologist.

What This Data Is (and Is Not)

The paper is very clear about what this dataset is for:

It IS: A powerful tool for exploration. It's great for spotting patterns, generating new ideas, or finding which species need a human expert to look at them next. It's a "first draft" of a global trait encyclopedia.
It IS NOT: A replacement for a scientist measuring a specific animal in a lab. You shouldn't cite a single answer from this dataset as an absolute, unchangeable fact without checking it yourself, especially for rare or weird species.

The Bottom Line

This paper presents a massive, cross-kingdom map of life's traits, built by a web-searching AI. It's like having a telescope that lets you see the general shape of the universe of life instantly, rather than having to walk to every star to measure it by hand. It's not perfect, but it gives scientists a starting point to ask better questions and find the most interesting species to study next.

Technical Summary: Cross-kingdom Phenotype Annotations for 35,856 Species

Problem Statement
Standardized organismal trait matrices are essential for comparative biology, macroecology, and conservation. However, assembling dense matrices with broad taxonomic coverage is hindered by the slow pace of manual curation, uneven taxonomic representation, and structured missingness driven by clade, geography, and research effort. While automated text mining and large language models (LLMs) have accelerated trait extraction from unstructured text, these approaches remain bounded by the availability and bias of the source corpus. Furthermore, cross-kingdom annotation is particularly challenging because traits as diverse as cell walls, venom, and seed dispersal must be represented within a single data framework. Existing resources often lack the uniformity required for broad-scale analysis, and LLM-generated data frequently suffers from "fluent errors" and sensitivity to prompts, necessitating rigorous validation and uncertainty quantification.

Methodology
The study presents LifeDive, a dataset generated by a web-search-enabled LLM embedded in a scripted annotation pipeline.

Organism Selection: The source taxonomy was the Catalogue of Life (2026-04-18 XR). A selection script prioritized Animalia, Plantae, and Fungi, aiming for family-level coverage. The process retained approximately 5% extinct species and favored accepted names and those with English common names. An initial set of 35,000 species was refined by an LLM-assisted review (using GPT 5.5 Pro) to include familiar organisms absent from the initial list, resulting in a final set of 35,856 species.
Annotation Pipeline: Species were evaluated on 196 phenotype questions covering cellular biology, anatomy, physiology, ecology, behavior, life history, reproduction, environmental tolerance, and human associations. The model used was Grok-4-1-fast-reasoning via the xAI API, configured with the web_search tool to retrieve information for obscure taxa.
Response Scale: Annotations utilized a five-level ordinal scale: 0 (Definitely No), 1 (Probably No), 2 (Unknown/not applicable), 3 (Probably Yes), and 4 (Definitely Yes). The prompt instructed the model to return code 2 when an answer could not be determined.
Data Processing:
- Parsing: Responses were validated for structure; misaligned outputs were discarded and replaced with Unknown values rather than silently accepted.
- Imputation: Missing values (code 2 and out-of-range) were imputed using random-forest classification (H2O cluster). The imputation treated other retained phenotype items as predictors to generate a completed matrix for downstream analysis, while preserving the raw matrix for provenance.
- Dimensionality Reduction: A 14-factor item-response-theory (IRT) model with promax rotation was applied to the completed matrix to derive phenotype dimensions (e.g., "Plant-like," "Aquatic," "Dangerous").

Key Contributions

LifeDive Dataset: A cross-kingdom dataset containing 7,027,776 ordinal phenotype annotations for 35,856 animal, plant, and fungal species.
Methodological Framework: A reproducible pipeline demonstrating how web-search-enabled LLMs can be used as structured annotators rather than mere text readers, incorporating explicit schemas, constrained outputs, and validation steps.
Data Layers: The release distinguishes between:
- Raw Matrix: Original model responses including "Unknown" codes.
- Completed Matrix: Imputed values for analyses requiring rectangular data.
- Derived Products: 14-factor loading matrices and unit-weighted phenotype scores.
Validation Suite: Comprehensive technical validation materials, including external benchmarks, cross-model audits, and latent-structure analyses.

Results and Validation

External Benchmarking: The dataset was compared against independent trait resources (e.g., PanTHERIA, EltonTraits, FishBase, AmphiBIO). Broad taxonomic indicators showed near-ceiling correlations (e.g., Animalia vs. cellulose: $r = -0.99$ ; Plantae vs. cellulose: $r = 0.99$ ; Mammalia vs. hair: $r = 0.99$ ). Specific ecological variables also showed strong agreement (e.g., amphibian larval stage vs. metamorphosis: $r = 0.90$ ).
Cross-Model Audit: A random sample of 200 species was audited using a different LLM (Claude Opus 4.7) and web searches. The audit estimated 95–97% directional agreement for the supplied matrix. Definitive codes (0 and 4) were found to be more accurate than hedged codes (1 and 3).
Latent Structure: Factor analysis yielded 14 interpretable dimensions (e.g., "Plant-like," "Fungus-like," "Insect-like," "Dangerous"). Visualizations confirmed that the scores aligned with biological expectations (e.g., woody plants occupying high "Plant-like" and "Durable" regions).
Phenotype Similarity: A network analysis of the 200-species subset showed that organisms clustered by phylum and functional analogy (e.g., trees with trees, frogs with frogs), with high within-group correlations (e.g., 0.877 for Ascomycota).

Significance and Claims
The paper positions LifeDive as a computational phenotype resource designed for broad and uniform coverage across the tree of life. The authors explicitly state that the data is not a replacement for primary measurements, expert-curated species records, or measured continuous traits.

Instead, the dataset is intended for:

Exploratory comparative analyses and hypothesis generation.
Benchmarking future annotation systems against a fixed set of organisms and questions.
Prioritizing uncertain cells for targeted expert review.
Generating features for macroecological modeling.

The authors emphasize that the data product must be judged by its documented provenance, visible uncertainty (distinguishing raw vs. imputed values), and independent technical checks. They caution that exact regeneration of raw annotations is not guaranteed due to the dynamic nature of LLMs and web retrieval, and they advise users to treat the data layers according to their specific tolerance for uncertainty.

Technical Summary: Cross-kingdom Phenotype Annotations for 35,856 Species

More like this