Methodology
How articles are found, selected, analyzed and published.
Scope
Onko-Gene follows the literature on the somatic genetics of solid tumors: molecular alterations, theranostic biomarkers, sequencing and molecular diagnostic techniques, and bioinformatics and artificial intelligence methods applied to oncology.
Each article is assigned to one or more editorial channels, each defined by a question:
- Clinical and theranostics: could the article change a prescription?
- Molecular biology and technologies: could it change laboratory practice?
- Bioinformatics and artificial intelligence: could it change an analysis pipeline?
Sources
Articles come from PubMed, queried every day. The queries search titles and abstracts and combine three conditions: a term for a solid tumor, a genomic or molecular term, and a filter specific to each channel. They are designed to leave out literature reviews, meta-analyses and studies of hematologic malignancies.
Selection
Since November 2026, each new article first receives a probability of publication, estimated by a language model (Gemini 3.6 Flash) from the title, the abstract and the journal name, following the site’s editorial policy. The model, its instructions and its settings are those evaluated on the curator’s past decisions, and they do not change. Only the articles whose probability reaches 0.42 are analyzed, then published. This threshold, chosen on the same decisions, gives roughly the curator’s pace: about ten articles a week.
Analysis
For each article retained, a language model reads the title and the abstract, never the full text, and produces in a single pass:
- the channels concerned;
- an assessment against twelve criteria rated on short scales: clinical impact, diagnostic impact, strength of evidence, novelty, sample size, journal quality, preprint status, therapeutic actionability, technological maturity, regulatory relevance, reproducibility and relevance across tumor types;
- a summary and a critical appraisal, in English and in French;
- the entities mentioned: genes, alterations, tumor types, therapies and keywords.
Ranking
For each channel, a composite score combines some of these criteria with a fixed formula, computed without a language model. It orders the articles of each channel in the weekly newsletter. Until September 2026, it also determined which articles were put forward for review.
Publication
Since November 2026, the selected articles are published without human review. Until September 2026, no article was published automatically: a curator examined every proposal, published or rejected it, and could correct the summary, the appraisal, the channels and the entities. Every entry states how the article was selected.
The articles selected during the week are published together on Tuesday, and the Wednesday newsletter carries them. Only published articles appear on the site; scores are not displayed.
Nomenclature
Tumor types are mapped to OncoTree (Memorial Sloan Kettering Cancer Center, version of October 3, 2025) when the match is exact or has been checked; otherwise the label taken from the article is kept. Gene symbols are set in italics, following HGNC nomenclature practice. References follow the Vancouver style.
Limitations
- The summary and the appraisal rely on the abstract alone. They do not replace reading the original article, linked from every entry.
- A language model can make mistakes or miss a nuance. Since November 2026, no human review corrects them before publication.
- The automatic selection also errs: on past decisions, about one in two articles kept at the 0.42 threshold had been rejected by the curator.
- The content is for information only. It is neither medical advice nor a treatment recommendation.