Gene Browser (/)

A paginated table of protein-coding genes. Its job is to get you to a gene and to make the annotation you are looking at explicit.

Model and Annotation are two different axes

This trips people up, so it is worth stating directly.

• Model picks a base predictor. It also sets the default annotation, because each model was trained against one — that mapping is the point, not an accident. openspliceai was trained on MANE GRCh38, so selecting it shows you MANE.
• Annotation overrides that. Browsing genes needs no model at all; it needs a GTF.

Annotation	Build	Protein-coding genes	Notes
MANE	GRCh38	19,288	One transcript per gene. Trains M1-S and M3-S.
Ensembl 112	GRCh38	20,089	All transcripts. Trains M2-S.
GENCODE v47	GRCh38	20,092	Near-superset of Ensembl (+136,858 sites genome-wide, mostly non-coding).
Ensembl 87	GRCh37	20,356	Legacy build, for the SpliceAI base model.

Different builds are not comparable

Selecting a GRCh37 annotation while a GRCh38 model is chosen shows a notice and disables the gene links. Coordinates from one build mean nothing to a model trained on the other, so the UI refuses to hand you through to a prediction rather than quietly reconciling them.

The first load of a new annotation parses its GTF (4.5–54 s), then caches to Parquet.

Search: name, ID, synonym, or description

Gene symbols are frequently not what a gene is known as. The ALS gene everyone calls TDP-43 is filed as TARDBP, and searching TDP returns three unrelated tyrosyl-DNA phosphodiesterases.

Search therefore covers four fields:

Query	Matches via	Result
TARDBP	symbol	TARDBP
TDP-43	synonym	TARDBP
TAR DNA binding	description	TARDBP
ALS10	synonym	TARDBP

Synonyms come from the RefSeq GFF (gene_synonym), which carries them for 16,890 of 19,288 MANE genes. The GTF has none, which is why they were invisible for so long. Ensembl and GENCODE GTFs also carry none, so they borrow the RefSeq map joined on gene name — a synonym belongs to the gene, not to the annotation listing it.

Results are relevance-ranked: exact symbol → exact synonym → symbol prefix → symbol substring → description-only. Without that, BRCA1 returns ten rows led by BRAP, because BRIP1, BARD1 and BABAM1 all mention BRCA1 in their descriptions.

Not found?

Literal substring matching only. TDP-43 works because it is a recorded synonym; a functional query like "genes involved in nonsense-mediated decay" will not. Semantic search over the agentic layer is on the roadmap, not built.

Columns

n_splice_sites is the count for that gene in the selected annotation, so it changes when you switch. TARDBP: 10 in MANE, 40 in Ensembl. That gap is the entire subject of the Genome View — those 30 extra sites are what M2-S exists to find.

Next

Click a gene name to open the Genome View.