PESTO

Prior Evidence Status from Text and Ontologies. Given a gene and a phenotype, PESTO says how much prior evidence links them, on four ordered levels:

Two branches answer independently, and are kept apart on purpose:

• Literature. PubMed is searched under the phenotype's synonyms and the gene's aliases, Claude Haiku 4.5 sorts every hit into four relevance bands, and Claude Opus 5 reads up to twenty abstracts and spreads 100 points over the four levels, citing the PMIDs it relied on.
• Open Targets. The gene's associated traits are fetched from the Open Targets Platform, the twenty nearest to the phenotype are ranked by an embedding model (BioLORD-2023), and Claude grades how each relates to the queried phenotype. The association score of a matching trait sets the level.

PESTO runs from the command line, or in a web page on your own computer with pesto browser: one pair or a whole table at a time, each pair opening on its verdict, the reasoning behind it, the articles read and links to gene databases.

This is the code of Large language model classifies prior evidence in gene–phenotype associations (Guez et al., 2026). The paper/ directory redraws every figure and table of the manuscript.

Install

PESTO is a command-line tool with a web interface, and pipx installs it in an environment of its own:

pipx install pesto-genetics

or, in a virtual environment you manage yourself:

python3 -m venv ~/venvs/pesto
~/venvs/pesto/bin/pip install pesto-genetics
~/venvs/pesto/bin/pesto --help

On Debian and Ubuntu a plain pip install outside a virtual environment is refused (externally-managed-environment); use one of the two above. On macOS, brew install pipx provides pipx. PESTO runs on Linux and macOS.

To update: pipx upgrade pesto-genetics.

The install is about 1.3 GB, nearly all of it PyTorch, which the Open Targets branch uses to rank traits locally. On Linux, pip's default index serves the CUDA build of torch; nothing here uses a GPU, and the CPU build is several hundred megabytes smaller (on macOS the default build is already the right one):

pipx install pesto-genetics --pip-args="--extra-index-url https://download.pytorch.org/whl/cpu"

The embedding model weights (439 MB) are downloaded on first use into ~/.cache/huggingface. pesto --download-models fetches them beforehand.

Without PyTorch

PyTorch serves one step only: ranking a gene's Open Targets traits by closeness to the phenotype, so that Claude grades the 20 nearest. Without it the install is about 200 MB:

pipx install pesto-genetics --pip-args="--no-deps"
pipx inject pesto-genetics anthropic pandas numpy requests python-dotenv

or, in a virtual environment:

python3 -m venv ~/venvs/pesto
~/venvs/pesto/bin/pip install anthropic pandas numpy requests python-dotenv
~/venvs/pesto/bin/pip install --no-deps pesto-genetics

PESTO notices that PyTorch is missing, says so, and reads every trait Open Targets associates with the gene instead of the 20 nearest (the same as --ot-encoder none). What that changes:

• The literature branch is unaffected: same searches, same reading, same verdicts.
• The Open Targets branch costs much more for well-studied genes, because Claude Opus grades every trait instead of 20. The cost grows with the number of traits Open Targets associates with the gene: a few cents for a gene with a few dozen, about $0.50 for one with several hundred (PCSK9 and familial hypercholesterolemia: 438 traits, $0.53 for this branch against $0.03 with PyTorch, and three minutes instead of one).
• Its verdicts can differ at the margin from the paper's, which used the BioLORD ranking. The grader sees every associated trait rather than the 20 nearest, so it can find a related trait the ranking would have left out, and it reads a list of several hundred less closely than a list of 20.

Installing torch and transformers afterwards restores the default.

In a web page

pesto browser

opens PESTO in your browser: run one pair, or drop a table of pairs and follow it as it runs; scroll through the answers, filter them by verdict, and open a pair to read its summary, the distribution over the four levels, the articles read (each opens in PubMed or Europe PMC), the Open Targets traits and how each relates to the phenotype, the cost, and links to gnomAD, GeneCards, Open Targets, OMIM, ClinVar, DECIPHER, the GWAS Catalog, GTEx and UniProt.

• Every table is a run of its own, kept in a folder under ~/.local/share/pesto/runs (change it in the settings, or with pesto browser --runs-dir DIR). The sidebar lists them; a run stopped or interrupted resumes where it was, without paying again for the pairs done.
• Runs made elsewhere open too: a folder written by pesto --bench (it holds pesto.tsv) opens with Open a run folder.
• Export as HTML saves a run as a single file that opens by double-clicking, with no installation: it can be read and shared, not run again.
• The API key is read from ANTHROPIC_API_KEY as on the command line. If it is not set, the page asks for it; the key is then kept in memory while pesto browser runs, never written to disk, and sent only to api.anthropic.com. The optional NCBI key can be pasted in the settings the same way, or read from NCBI_API_KEY.

The page is served on 127.0.0.1 only, and every request must carry the token in the address pesto browser opens, so no other site open in the browser can start a run. Before a table runs, the page shows its estimated cost.

Your Anthropic API key

PESTO calls Claude through the Anthropic API, with your own key, billed to your account. Create one at console.anthropic.com, then:

export ANTHROPIC_API_KEY=sk-ant-...

or put the same line, without export, in ~/.local/share/pesto/.env.

The key stays on your machine. PESTO never writes it to disk, never prints it, and never stores it in a saved run. It is read in one place, src/pesto/credentials.py, and handed only to the official anthropic client in src/pesto/services/llm_service.py, which sends it to api.anthropic.com and nowhere else. Those two files are short; read them to check.

An NCBI key (NCBI_API_KEY, optional) raises the PubMed rate limit from 3 to 10 requests a second and is handled the same way.

On the command line

pesto --gene DDX41 --phenotype "myelodysplastic syndrome"

prints the literature verdict, the Open Targets verdict, how many articles were found and read, and what the pair cost. --json returns everything, including the probability distribution, the justification and the supporting PMIDs.

A table of pairs (tab-separated, columns gene and phenotype) runs in parallel:

pesto --bench pairs.tsv --workers 10      # writes pesto.tsv and cost.tsv beside it

pesto --help lists the rest.

As of a past date

pesto --gene DDX41 --phenotype "myelodysplastic syndrome" --max-date 2014

asks what the literature said at the end of 2014. DATE is YYYY, YYYY/MM or YYYY/MM/DD, and a year or a month stands for its last day. Every PubMed search, and the Europe PMC search of --fulltext, is restricted to articles published by then, and any article whose displayed year is later is dropped as well.

The model that reads the abstracts is also told, in a note added before them, that they are the literature as it stood on that date and that it should judge from them alone, drawing on nothing else it knows (prompts/max_date_note.txt). --no-date-note sends the standard prompt instead.

Three things it does not do, and PESTO says so when it starts:

• Open Targets is not run. Its API serves only the current release, whose association scores cannot be restricted to a date, so the Open Targets columns are left empty (ot_channel reads not run: --max-date).
• The reading model is not dated. Claude may know of work published after the date. The note asks it not to use that knowledge, which narrows the leak without closing it.
• The knowledge arm searches nothing, so the date has nothing to restrict, and it is sent no note.

The date is part of the pipeline's fingerprint, so a cached run, or a row kept by --resume, is reused only for the same date; runs without --max-date keep their fingerprint. Tables carry it in a max_date column.

Notes to the reader

--note "..." adds a sentence to the reading prompt, on a line of its own before the articles, and can be repeated:

pesto --gene DDX41 --phenotype "myelodysplastic syndrome" \
      --note "Somatic variants in tumours do not count as evidence."

With --max-date, the date note comes first and yours follow in the order given. Notes are part of the fingerprint, so a run is reused only for the same notes, and tables carry them in a notes column. A run with notes no longer reads the prompt the paper's numbers were measured with. The knowledge arm refuses notes, having no articles to put them before.

Pipelines

A pipeline ("arm") is declared in src/pesto/flow/arms.py as a list of blocks. Its fingerprint hashes every block's parameters and prompts, and every saved run carries it, so a result says exactly which pipeline produced it.

By default PESTO asks Claude whether the phenotype is a disease, and runs:

• abstracts for a disease: PESTO as in the paper.
• abstracts-trait otherwise: the same pipeline with prompts that do not assume a disease, so replicated cohort and GWAS evidence can reach Established.

With --fulltext, it runs fulltext or fulltext-trait instead. These add a Europe PMC search for articles that report the gene only in their body (a results table, a supplementary file) and append the passages naming it for the reader. PubMed cannot return those articles, and the deciding paper is often one of them. On 20 GWAS associations whose deciding article names the gene only in its body, abstracts placed 1 at Existing or above and fulltext 18, and both kept all 20 negative controls at Novel or Hypothesized (Supplementary Figure 3). fulltext was not used for the analyses of the paper, and fulltext-trait has not been benchmarked.

The other arms reproduce the paper's comparisons: titles (titles only), knowledge (the model's own knowledge, no search), and abstracts-bare, abstracts-open and abstracts-score (reading-prompt ablations).

Cost

At Anthropic list prices, the literature branch costs about $0.07–0.11 a pair with abstracts (Extended Data Fig. 2 and Supplementary Figure 3), and 2–3% more with fulltext; the Open Targets branch about $0.03. Well-studied genes cost more. A pair already answered by the same pipeline is read from the cache and costs nothing. With --bench, the cost of every pair is written to cost.tsv as it runs.

Where things are written

Runs are saved under ~/.local/share/pesto (or $PESTO_PROJECT_ROOT), one JSON per pair with the corpus, the prompts and the model's answer.

Reproducing the paper

See paper/README.md. Every figure and table is redrawn from the tables in paper/results/ without calling any model. The saved model answers behind them are on Zenodo: doi:10.5281/zenodo.23230222.

Tests

Free, no model calls, no network:

python3 tests/replay.py           # fingerprints, prompts, selection rules, parsing
python3 tests/script_imports.py   # every paper script still imports

Citation

Guez J, Auwerx C, Lu W, Satterstrom FK, Berkowitz J, Fu JM, Betancur C, Talkowski ME, Daly MJ, Karczewski KJ. Large language model classifies prior evidence in gene–phenotype associations. 2026. Code and data: doi:10.5281/zenodo.23230222.

Licence

Code: MIT. Tables and model answers produced for the paper: CC-BY 4.0. Third-party content they include keeps the terms of its source.