LLMarkersDB

From papers to grounded marker evidence

mrkr reads each paper and uses a language model to return an exact source passage, a normalized marker statement, and normalized labels for the cell type, marker genes, comparison groups, tissues, and organism. The model does not assign database identifiers.

Deterministic checks confirm that quoted text and explicit labels occur in the paper. A separate grounding step maps accepted genes to Ensembl, cell types and comparison groups to the Cell Ontology, tissues to UBERON, and organisms to NCBI Taxonomy. Cell Ontology candidates are verified against a pinned ontology release and accepted only when the normalized label is a canonical label or exact synonym. The validated record is stored in .onto; this SQLite database is a derived representation of those records.

What the site searches

Search matches normalized target labels, verified Cell Ontology identifiers, gene symbols, and Ensembl identifiers. A query containing several recognized genes ranks aggregated cell type marker panels against the gene set. Search runs entirely in the browser against llmarkers.sqlite.

Search results group records by verified Cell Ontology identifier when one is available and otherwise by normalized cell type label. Each card shows the union marker panel. The number beside each gene is its count of supporting normalized statements. Selecting a gene shows those statements, their papers, and the exact source text.

The Evidence tab shows one row for each cell type–gene pair in a validated marker statement. Each row shows the paper, target cell type, marker direction, reported comparison set, tissue, and normalized statement.

About LLMarkers

LLMarkersDB contains source-grounded marker evidence from 979 bioRxiv and Human Cell Atlas-linked papers. It connects author labels across papers with stable identifiers while retaining the text that supports each connection.

Stable identifiers improve label reuse, but they do not replace experimental context. Comparison groups determine whether a gene distinguishes a target population. LLMarkersDB therefore preserves reported comparisons when authors provide them and makes missing comparisons visible.

Source code, validated artifacts, formalization, and reproduction commands are available in the llmarkers repository. The extraction and validation tool is maintained in the mrkr repository.