About oncolabels

oncolabels lists the indications the EU has approved for cancer drugs and the biomarker result each one names. It is a non-commercial reference for education and research, written for pathologists, oncologists and researchers.

What it covers

Every antineoplastic or endocrine cancer drug the EU authorises centrally (ATC L01 or L02, or EMA's group of that name while the code is pending), apart from classic chemotherapy, plus the immunomodulators and interferons given against cancer and the radioligands: 187 drugs with 604 indications, from EMA data checked on 5 Nov 2026. Each entry is one approved indication with one biomarker condition, taken from section 4.1 of the summary of product characteristics (SmPC). Biosimilars and generics are listed on the page of the reference drug.

How the entries are made

For each drug, oncolabels downloads the current SmPC from the EMA website and reads section 4.1. EMA's medicines data is updated twice a day but can lag behind the PDF, so the PDF is the source.

Two language models from different vendors, Gemini 3.8 Flash and GPT-6 Luna, read the section separately. Each splits it into single indications and fills in the tumour, setting, line of therapy, biomarker, cut-off, regimen and conditions. A third model, Jev 1.14, classifies the categorical fields again and removes indications outside oncology. Where the readings differ, Claude Opus 5.5 reads the label, decides and gives its reason. If the label is still ambiguous after that, the entry keeps the first reading and gets an amber dot.

The target of each drug (PD-1, CD79b, KRAS G12C) comes from the mechanism of action in section 5.1 of the same SmPC, read the same way by two models, with a third deciding where they differ.

No person has checked the entries yet. Each entry quotes the label sentence it comes from and links to the SmPC in English and German, and its record shows what each model read and decided. The methods pages list the extraction rules and every decision the models made.

How the search reads a query

Your words are matched against a vocabulary of tumour types, biomarkers, drugs and treatment settings, with common abbreviations and misspellings. A score such as “TPS 30%” counts as a patient result. Each term the search recognises appears as a chip under the search box; remove a chip to widen the search.

Reading the results

  • Line of therapy follows the label. A line filter also shows approvals with no line restriction, and later lines include labels that say “after prior therapy”.
  • Histology follows the label too. Where a label limits an approval to one histology, such as squamous cell carcinoma of the oesophagus or non-squamous NSCLC, the Conditions column says so, and the tumour's page lets you pick a histology. Approvals for any histology stay under every choice. Adenocarcinoma of the gastro-oesophageal junction appears under gastric and oesophageal cancer.
  • Advanced disease covers locally advanced unresectable and metastatic disease, as in ESMO guidelines, so an approval for advanced disease appears under both.
  • Patient results. Type a score such as “TPS 30%” or “CPS 5” with a tumour, and approvals whose cut-off the score does not meet are hidden. Approvals scored with another system stay in the list and are marked.
  • Tumour-agnostic approvals appear with every solid tumour, in a section of their own.
  • What to test groups the biomarkers by how pathology labs test them: by immunohistochemistry, by molecular testing, or by either (alternative methods, or IHC screening confirmed by molecular testing). The label usually asks only for a validated test, so the routes are set from pathology practice.

API and MCP

AI assistants such as Claude and ChatGPT can look up the same data through an MCP server, and scripts through a JSON API. Set up the API or the MCP server.

Limits

oncolabels reads section 4.1 only. It does not cover assays named elsewhere in the SmPC, national reimbursement decisions such as those of the G-BA, FDA approvals or guideline recommendations. It gives no medical advice, and the current SmPC takes precedence over anything shown here.

Privacy

Your browser downloads the dataset once, and searches run on your device. Only a search with words the search does not know is sent, through this site's server, to an AI model (Jev 1.14 by TypeSafe, through OpenRouter, a US service), which maps the words to a tumour, biomarker, line or setting. Its readings are marked “read by AI” and can be removed. Do not type patient identifiers.

Licence

The data is free to reuse under CC BY 4.0. Credit oncolabels and the European Medicines Agency. Download it as JSON or CSV.