BiomedCAT: Biomedical Context Analysis Tool for Supporting Mechanistic Hypothesis Generation via User-defined Contextual Subgraph Extraction

Researchers formulate mechanistic hypotheses in slides and figures long before they write them as text, yet knowledge-graph (KG) tools expect a query in the vocabulary of the graph. We present BiomedCAT, an open-source workflow that turns biomedical slides into a subgraph of the RTX-KG2c knowledge graph shaped by the user's research profile, on a laptop and without sending the documents out. A small vision-language model describes each slide; a 4-billion-parameter agent types the terms into the classes of the Biolink model; an existence gate removes invented entities; identifiers are retrieved from Translator services or from a local table and selected, never generated, by a judge model; and a profile-weighted two-hop expansion builds the context graph. On a synthetic Biolink typing benchmark whose labels were audited by an expert, the local agent reaches 0.71 accuracy on 517 non-ambiguous mentions, 4 points below Claude Opus 5 (0.75), level with DeepSeek-V4-Flash (0.69) and above open-weight models of 24 and 31 billion parameters; the audit also showed that 55\% of the generated labels were wrong or ambiguous. On expert-annotated slide decks, the reading stage recovers 81 to 94\% of the explicit terms of the English deck, less on French decks, and linking assigns the expert identifier to 65\% of the found terms online and 26\% in the fully local mode. A one-slide run takes 33 to 42 minutes, with peaks of 9.8~GB of system RAM above idle and 5.3~GB of GPU memory. In a facioscapulohumeral muscular dystrophy (FSHD) case, three slides that each draw one pathway from DUX4 at a different biological scale give graphs that overlap by 7 to 15\% of their nodes, each recovering its own pathway's expected nodes, and expose sourced paths from DUX4, such as a DUX4--HMGB1 interaction from IntAct. Of 21 compounds tested in FSHD, seven, including every agent that targets DUX4, are absent from RTX-KG2c, which bounds what a graph-based approach can propose. BiomedCAT lets a researcher set a working hypothesis against curated knowledge, with every edge traceable to its source.

Authors

Institutions

Publication Details

Journal
bioRxiv (Cold Spring Harbor Laboratory)
Published
2026-11-18
DOI
https://doi.org/10.64898/2026.11.11.766070
Primary Topic
Biomedical Text Mining and Ontologies
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

BiomedCAT: Biomedical Context Analysis Tool for Supporting Mechanistic Hypothesis Generation via User-defined Contextual Subgraph Extraction

Sandrine Muller, Laurent Torlay, Julie Dumonceaux, Christophe Hourdé et al.
2 links
bioRxiv (Cold Spring Harbor Laboratory)
Biomedical Text Mining and Ontologies
preprint

BiomedCAT: Biomedical Context Analysis Tool for Supporting Mechanistic Hypothesis Generation via User-defined Contextual Subgraph Extraction

Sandrine Muller, Laurent Torlay, Julie Dumonceaux, Christophe Hourdé, Salah Eddine Chahma, Lucca Hoogenbosch
preprint en

Abstract

Researchers formulate mechanistic hypotheses in slides and figures long before they write them as text, yet knowledge-graph (KG) tools expect a query in the vocabulary of the graph. We present BiomedCAT, an open-source workflow that turns biomedical slides into a subgraph of the RTX-KG2c knowledge graph shaped by the user's research profile, on a laptop and without sending the documents out. A small vision-language model describes each slide; a 4-billion-parameter agent types the terms into the classes of the Biolink model; an existence gate removes invented entities; identifiers are retrieved from Translator services or from a local table and selected, never generated, by a judge model; and a profile-weighted two-hop expansion builds the context graph. On a synthetic Biolink typing benchmark whose labels were audited by an expert, the local agent reaches 0.71 accuracy on 517 non-ambiguous mentions, 4 points below Claude Opus 5 (0.75), level with DeepSeek-V4-Flash (0.69) and above open-weight models of 24 and 31 billion parameters; the audit also showed that 55\% of the generated labels were wrong or ambiguous. On expert-annotated slide decks, the reading stage recovers 81 to 94\% of the explicit terms of the English deck, less on French decks, and linking assigns the expert identifier to 65\% of the found terms online and 26\% in the fully local mode. A one-slide run takes 33 to 42 minutes, with peaks of 9.8~GB of system RAM above idle and 5.3~GB of GPU memory. In a facioscapulohumeral muscular dystrophy (FSHD) case, three slides that each draw one pathway from DUX4 at a different biological scale give graphs that overlap by 7 to 15\% of their nodes, each recovering its own pathway's expected nodes, and expose sourced paths from DUX4, such as a DUX4--HMGB1 interaction from IntAct. Of 21 compounds tested in FSHD, seven, including every agent that targets DUX4, are absent from RTX-KG2c, which bounds what a graph-based approach can propose. BiomedCAT lets a researcher set a working hypothesis against curated knowledge, with every edge traceable to its source.

bioRxiv (Cold Spring Harbor Laboratory)
Université Claude Bernard Lyon 1 (FR), Institut polytechnique de Grenoble (FR), Centre National de la Recherche Scientifique (FR), Great Ormond Street Hospital (GB), Laboratoire d'Informatique de Grenoble (FR), Laboratoire Interuniversitaire de Biologie de la Motricité (FR), NIHR Great Ormond Street Hospital Biomedical Research Centre (GB), University College London (GB), Université Savoie Mont Blanc (FR), Université Grenoble Alpes (FR)
Université Grenoble Alpes, Agence Nationale de la Recherche, Centre National de la Recherche Scientifique, National Institutes of Health, National Center for Advancing Translational Sciences
Biomedical Text Mining and Ontologies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon