Human genome browser¶

Open the browser to explore precomputed GRCh38 SNV delta scores with native IGV.js navigation and an OpenSpliceAI exact-allele inspector.

The bundled demonstration contains ten real scored regions from the final r10 and r13 collections. Its reference excerpts, gene features and scores are self-contained. Regions outside the demonstration are explicitly unavailable. The completed source collections and their genome-wide browser packaging are separate deliverables; full online data require a configured indexed-data host.

Scored demonstration regions¶

The demonstration contains 67,579 source occurrences. The Scored examples buttons navigate directly to published windows. Coordinates below are 1-based inclusive; these windows are examples, not whole-gene score releases.

Table 1 Published GRCh38 score windows¶
Example	Interval
OR4F5 (+)	chr1:68,991–70,110
OR4F16 (−)	chr1:685,501–686,900
SAMD11 (+)	chr1:925,401–930,500
Alternate-contig audit	chr2_KI270773v1_alt:18,501–19,400
Strong delta-score example	chr2:241,871,201–241,872,225
Matched r10/r13 difference	chr2:241,869,354–241,870,378
CFTR (+), splice junction	chr7:117,503,340–117,505,387
HBB (−), splice junction	chr11:5,225,553–5,227,600
BRCA1 (−), splice junction	chr17:43,122,993–43,125,040
BRCA2 (+), splice junction	chr13:32,315,504–32,317,551

The coverage notice distinguishes a published interval, partial overlap and unavailable scores. Partial views offer a button to focus an overlapping scored example. Gene searches can extend beyond the demonstration; they do not imply that scores for the entire gene are published. Exact and summary CSV downloads are disabled outside published score windows. The mitochondrial sequence is a reference-search example and has no demonstration score window.

Navigate and inspect¶

Enter a scoring-gene symbol, a 1-based inclusive region such as chr1:925851-926130, or an exact allele such as chr1:69091:A>G. Drag a native track to pan continuously. Drag the ruler to zoom to a region; hold Shift while dragging the ruler to mark a region of interest. Native track menus control presentation, scaling and height; drag handles reorder tracks. The controls also support chromosome navigation, zoom, previous/next genes, multiple genomic panes, and explicit selection of the inspector’s active pane.

The Exact ALT matrix and matched differences panel follows the native IGV columns beneath every displayed region. Columns retain their genomic alignment while dragging, zooming and resizing, including on high-DPI displays. Select a column to inspect that pane and base. The inspector, comparison statistics and CSV exports follow the selected pane. Newly exposed bases are labelled as loading until their exact records arrive; absent predictions remain gaps.

The exact table displays 100 alleles per page. Pagination does not truncate statistics or CSV exports. Select an allele to inspect all original gene annotations, source occurrences, five-decimal OpenSpliceAI values and signed offsets. Native overview tracks show maxima across REF-matching alternate alleles and genes; ALT, gene and threshold filters apply to the exact panels.

Save and restore a session to preserve the immutable manifest hash, model, view settings, loci, track presentation and marked regions. Shared links pin the manifest URL and content hash. BED regions can be imported and exported. Additional local or URL tracks use IGV’s supported formats and corresponding indexes. Local files remain on the user’s computer and must be reselected when restoring a session on another page or device.

Interpret the measurements¶

DS_AG / DS_AL: acceptor gain / loss variant delta scores.
DS_DG / DS_DL: donor gain / loss variant delta scores.
DP: original signed offset. The affected genomic coordinate is VCF POS + DP on both gene strands. A zero DS does not identify an affected site; its original DP remains available in the inspector and CSV.
Coverage: REF-matching input and predicted ALT counts. Absent predictions, measured zero, REF mismatches, positions outside the input collection and regions outside the demonstration remain distinct.
r13 minus r10: matched allele/gene comparisons. Conflicting annotations and REF mismatches are excluded from ordinary comparison statistics.

Both final source collections contain 3,433,384,833 source occurrences. Of 3,339,752,304 distinct canonical input SNVs, 3,271,090,620 have predictions per model (approximately 97.94%). The remaining 68,661,684 do not have a prediction annotation. Completing processing does not turn absent predictions into zeros. The source collection does not contain every possible human SNV.

r10 and r13 are individual MANE-trained model seeds, with 10,000 nt flanking size, event search distance 50 and M1 masking. They are not a model ensemble. The scoring genes come from the pinned GENCODE v24 canonical annotation. The original masked SpliceAI baseline has two-decimal source precision and a different annotation history. Model disagreement is not calibrated uncertainty.

Wide views use indexed annotation maxima. Exact inspection and CSV retain the lossless source scores, offsets, source identities and duplicate occurrences. The evidence panel distinguishes legacy data from receipt-bound provenance; a model configuration label does not establish receipt-bound identity for historical files. Reference, annotation, model and dataset hashes are available in the methods panel.

Export and sequence search¶

PNG and vector SVG export native IGV figures, including whole-genome overviews. Exact CSV includes all matching original annotations and explicitly named 0-based half-open and 1-based coordinate columns. Missing values are blank; recorded zeros remain numeric zero. Summary CSV is a separate export of bins. Exports require stable current-view data.

IUPAC sequence search supports both strands and cancellation. Visible-region search uses published reference excerpts; chromosome and whole-reference search require the corresponding static search indexes. The demonstration includes the complete mitochondrial reference and its search index. Unpublished indexes produce an explicit unavailable result rather than zero matches.

Run and review locally¶

The browser performs no model inference. With Node.js 22.12 or later:

cd docs/genome-browser
npm ci
npm test
npm run build
npm run dev

Open http://127.0.0.1:4173/OpenSpliceAI/genome/. To serve a verified full snapshot and the built application from one local origin, run from the repo root:

python -m tools.genome_browser serve SNAPSHOT_DIR \
    --browser docs/genome-browser/dist/genome --port 8765

Then open http://127.0.0.1:8765/. The data server serves the snapshot’s manifest-listed files; preparation databases and private configuration are excluded. The legacy viewer preserves version-1 snapshot links.

Core release checks¶

npm test runs 22 core checks for score precision, missing values, complete queries, published coverage and native coordinate transforms. After npm run build, npm run audit runs seven production-browser workflows in each of Chromium, Firefox and WebKit. These cover navigation, selection, exports, sessions, data errors and usable phone controls. Two alignment scenarios compare drawn matrix and difference columns with native one-base markers through navigation, resizing, high DPI and multiple regions, with a one-physical-pixel tolerance. Install their runtimes with npx playwright install --with-deps chromium firefox webkit first.

The complete frontend and visual suites remain available as npm run test:all and npm run audit:all. The separate Genome browser extended audit workflow runs those checks and the full data-preparation tests on request. Pages deployment requires core functionality and public-file integrity; broad appearance and document scroll-area audits run separately.

Prepare and host full data¶

CPU-only preparation reads the accepted, content-hashed source VCFs. Packed blocks, native BigWigs, indexed genes and reference/search artifacts support bounded regional access without millions of tile files. Configurations and builder code are frozen for scheduled runs. Resume authenticates completed packs and their source identities; full publication requires a final verified release, including source census reconciliation and reference content proof.

The large data stay outside Git and GitHub Pages. A future HTTPS data endpoint must provide correct HTTP 206 byte ranges, expose Content-Range through CORS when cross-origin, and preserve indexed bytes without HTTP content compression. Export a verified file inventory and validate the actual endpoint before activating its catalog. Publishing is a separate action from local review. See tools/genome_browser/README.md for preparation and hosting commands.