We have recently released several versions of the Cell Line Verifier and integrated several OpenAI models into our workflow. Each version can be combined with one or more models. As these combinations grow more complex, this guide breaks down their characteristics and recommended uses.

For any question about the API, please see the API documentation.

Cell Line Verifier versions

The versions currently available are listed below:

Version	Description
v1	First released version. Extracts the cell lines cited in the article text and checks each one against Cellosaurus. Only supports gpt-4o.
v1.1	Enhanced version of v1, recommended for most uses. It also reads complicated tables, and is stricter when linking a self-derived cell line to its likely parent, which avoids wrong matches. Supports gpt-6-luna (default) and gpt-4o.
v1.1.img	Same core functioning as v1.1, but also reads the article's figures: each image is described in the light of its caption, so cell lines shown only in images (e.g. a screening panel in a graph) are found too. Slower (up to about a minute per article) and uses several times more tokens on illustrated articles. Only supports gpt-6-luna. A name read in a figure but absent from the article text is tagged figure only: check the figure, it may be a misreading.

This information is valid as of 7 October 2026. Some of these versions may change or be withdrawn in the future.

OpenAI models

The OpenAI models currently integrated into our pipeline are listed below:

Model	Description
gpt-4o	Older-generation model, generally more consistent from one run to the next. A good choice when repeatability matters more than performance. It is also significantly faster, about 3 times faster than gpt-6-luna in our tests, which can make it worth using when speed matters. However, it is much more expensive, and its daily quota is 10 times smaller than that of gpt-6-luna.
gpt-6-luna	Powerful new-generation model. Its results can vary slightly between two runs of the same article (about 1–2% of cell lines in our tests). It is very cheap, roughly 10 times cheaper than gpt-4o. We recommend it when you have a large number of articles to process.

These models were tested and assessed on the OpenAI platform as of 7 October 2026. Some of them may change or be withdrawn in the future. Please note that this tool is only possible thanks to OpenAI's data-sharing program, which provides free daily tokens in exchange for sharing data. We warmly thank OpenAI for it.

General notes for users

We recommend sticking to gpt-6-luna and v1.1 for most day-to-day use. Only switch to gpt-4o or v1 for specific reproducibility or speed needs. Finally, avoid v1.1.img unless you specifically want to extract cell lines from figures and intend to screen only a handful of papers.

If a result surprises you, don't hesitate to rerun the analysis or to switch model or version: LLMs have some intrinsic variability, which can occasionally lead to unexpected results.

The Cell Line Verifier API is the easiest way to automate cell line extraction from PDFs within your own workflow. The daily usage limit is indexed to that of your online CLV account, which can be increased upon request. Please note that this service is running on the free-tier data sharing plan offered by OpenAI. So please do not input any sensitive data, or anything that you're not worth sharing with OpenAI.

To get started, click on Generate API key in your account settings and follow the instructions below.

Making a request

The only endpoint intended for external use is api/analyse, making the base URL:

https://cell-line-verifier.com/api/analyse

You can test the endpoint and your API key with a curl request. Send your key as a Bearer token in the Authorization header, and upload your PDF as multipart/form-data:

curl https://cell-line-verifier.com/api/analyse -H "Authorization: Bearer YOUR_API_KEY" -F "pdfs=@/path/to/paper.pdf"

Please note that only one PDF can be sent per request. You can send several requests in parallel: the server analyses up to 5 PDFs at once (all users combined), and the others wait in a queue and are processed in order.

By default the latest pipeline (v1.1) is used. To pick a version explicitly, add a tool_version form field (v1.1, v1 or v1.1.img), e.g. -F "tool_version=v1". v1.1.img also reads the article's figures (with gpt-6-luna only): it finds cell lines shown only in images, but takes longer and uses more tokens.

The extraction model can be chosen the same way with a model form field: gpt-6-luna (default) or gpt-4o, e.g. -F "model=gpt-4o". v1 only runs on gpt-4o, which is used automatically when no model is given. Please note that gpt-6-luna results can vary slightly between two runs of the same file.

Response format

On success, every request returns the same object:

{
  "tool_version": "v1.1",
  "model": "gpt-6-luna",
  "filename": "test_paper_1.pdf",
  "doi": "10.1186/s12935-023-03077-4",
  "extracted": ["AGS", "HT-22", "MKN45", "MKN47", "MNK45", "SGC-7901"],
  "verified": [["AGS", "CVCL_0139"], ["HT-22", "CVCL_0321"], ["MKN45", "CVCL_0434"]],
  "problematic": [["SGC-7901", "CVCL_0520"]],
  "proprietary": [["SKBR-3/Tr", "CVCL_0033"], ["MHCC97/SR", "CVCL_4971"]],
  "unverifiable": ["MKN47", "MNK45"],
  "figure_only": [],
  "counts": {"AGS": 87, "MKN45": 79, "MNK45": 7, "SGC-7901": 4, "SKBR-3/Tr": 2, "HT-22": 1, "MKN47": 1, "MHCC97/SR": 1},
  "predictions": {"MNK45": "recurring_mistake", "MKN47": "punctual_mistake"}
}

Field	Description
tool_version	The extraction pipeline version used for this file: v1.1 (default, recommended), v1 (previous version) or v1.1.img (v1.1 plus the figures)
model	The model used for extraction: gpt-6-luna (default) or gpt-4o
filename	The name of the uploaded file
doi	The article's DOI found in the PDF (e.g. 10.1186/s12935-023-03077-4), or null if none was found
extracted	Every cell line name found in the text, before verification
verified	[name, CVCL accession] pairs confirmed against Cellosaurus
problematic	[name, CVCL accession] pairs flagged as problematic (e.g. contaminated, misidentified)
proprietary	[name, parent CVCL accession] pairs for lines identified as proprietary/self-derived, linked to their likely parent line
unverifiable	Names that could not be matched against Cellosaurus
figure_only	v1.1.img only: the extracted names found in a figure but nowhere in the article text. They may be misreadings of the figure, so check it before relying on them.
counts	Number of times each name is written in the article text (references excluded), keyed by the name as listed above. Spellings of the same line are counted together; names found only in a figure are absent.
predictions	For each non-verifiable name written in the text: independent_use (used as a cell line of its own, e.g. listed alongside the real line it resembles), recurring_mistake (written more than once in place of one of the article's real lines) or punctual_mistake (once) or uncertain (looks like a mistake, but the name has 3 characters or fewer and is written at most twice: too short to tell). An independent use can be an unregistered line developed by a group: it calls for a check, not a verdict.

Please note that cell lines classified as proprietary or non-verifiable might require further verifications as this tool can make mistakes.

Python code example

Here's a Python example. Wrap this in a loop to process multiple files, or run up to 5 calls in parallel (e.g. with a thread pool of 5 workers).

import os
import requests

API_KEY = os.environ["CLV_API_KEY"]

def analyse_pdf(path, tool_version="v1.1", model="gpt-6-luna"):
    with open(path, "rb") as f:
        r = requests.post("https://cell-line-verifier.com/api/analyse",
            headers={"Authorization": f"Bearer {API_KEY}"},
            files={"pdfs": (os.path.basename(path), f, "application/pdf")},
            data={"tool_version": tool_version, "model": model},
            timeout=30,)
    r.raise_for_status()
    return r.json()

print(analyse_pdf("test_paper_1.pdf"))

The same process can be repeated in the programming language of your choice.