Genome View (/genome/{gene})¶
The main workspace: per-position splice-site probabilities for one gene, optionally with a meta model overlaid on its base model.
The four track bands¶
| Band | Shows |
|---|---|
| Donor Prob | per-position P(donor), base and meta |
| Acceptor Prob | per-position P(acceptor) |
| Classification | TP / FP / FN markers at the current threshold; meta as open diamonds |
| Annotations | stacked ground-truth rows — see below |
Genes over 10,000 positions are downsampled for rendering, but peak-preserving: every position above a small probability floor is kept, then evenly-spaced background fills the budget. Naive every-Nth slicing would drop the sharp 1–2 position spikes that are the signal.
Annotation tracks¶
Four rows, toggled by the checkbox panel below the chart (or by clicking a chart legend entry):
| Row | Default | Meaning |
|---|---|---|
| MANE | on | Canonical reference. What the base model was trained on. |
| Ens∖MANE | on | The delta set — alternative sites MANE does not contain. What M2-S exists to find. |
| Ensembl | on | All Ensembl sites. Visually contains MANE ∪ delta. |
| Gc∖Ens | off | GENCODE's additions. Essentially always 0 for protein-coding genes, so shown greyed. |
A row label reads Ens∖MANE (10) where the count is sites inside the scored window.
Sites outside the gene span cannot be scored
Prediction runs over the gene's span in the model's annotation. Ensembl genes are often longer: TARDBP is 18,184 bp in Ensembl against 12,838 in MANE, so 8 of its 30 alternative sites fall outside the window entirely. They are excluded from the denominator and clipped from the tracks, with the count stated under the badges. Counting them as misses would understate every model equally and for no reason.
The two thresholds¶
A splice model's raw output is a probability per position; a threshold turns that into a call. The base and meta models have very different score distributions, so a single shared cutoff always scores one of them at the wrong operating point.
- LINK checked (default) — one cutoff for both. The first slider is labelled
Threshold. - LINK unchecked — two independent cutoffs. The first slider relabels to
Base threshold.
Find F1-optimal sweeps this gene and reports both models' optima, each with its own Use button. Applying a meta threshold that differs from base unchecks LINK for you.
The panel reports two blocks, and names the truth set for each — read that label before quoting either:
| Block | Scored against |
|---|---|
| This gene (post-hoc) | the base model's annotation, canonical transcript only (MANE for openspliceai) |
| Held-out evaluation | the meta model's own eval — Ensembl, all annotated sites, 14,724 test genes |
Neither is the delta set. See Reading the numbers.
Reading base vs meta — worked example¶
Open /genome/TARDBP, base openspliceai, meta M2-S, threshold 0.5.
Ground truth auto-resolves
With a meta overlay active, TP/FP/FN is scored against that model's training annotation —
Ensembl for M2-S — shown as a vs ensembl tag beside the counts. The Ground truth selector
overrides it. The walkthrough below uses truth=mane to show why the default changed; set the
selector to MANE to reproduce it.
Scored on MANE, the meta model looks bad:
| TP | FP | FN | |
|---|---|---|---|
| Base | 10 | 0 | 0 |
| Meta | 10 | 39 | 0 |
Same true positives, 39 extra false positives. On this reading the meta layer hurts.
Now look at the Alt-sites badge: base 0/22 → meta 20/22. Of the 39 "false positives",
20 sit exactly on splice sites Ensembl annotates. They are false only because the ground truth
here is MANE, which lists one transcript per gene.
Raise the threshold to 0.9:
| thr | base TP/FP/FN | meta TP/FP/FN | meta FPs that are annotated alt sites | alt recall base → meta |
|---|---|---|---|---|
| 0.5 | 10/0/0 | 10/39/0 | 20 / 39 | 0/22 → 20/22 |
| 0.9 | 9/0/1 | 10/19/0 | 19 / 19 — all of them | 0/22 → 19/22 |
| 0.99 | 7/0/3 | 10/13/0 | 13 / 13 | 0/22 → 13/22 |
At 0.9 every meta false positive is a real annotated splice site. The unexplained ones at 0.5 were an operating-point artifact, not bad predictions. Meanwhile the base model has begun losing canonical sites — 9 of 10, then 7.
The summary: base finds 10 sites and misses 22; meta finds the same 10, plus 19 of the 22, with no unexplained false positives.
Now switch Ground truth to Ensembl (all) — the default when M2-S is overlaid, and the
annotation it was actually trained on. Same predictions, threshold 0.9:
| TP | FP | FN | |
|---|---|---|---|
| Base | 9 | 0 | 23 |
| Meta | 29 | 0 | 3 |
29 of 32 sites with zero false positives, against base's 9 of 32. The 19 "false positives" from the MANE view were never errors — they were the yardstick refusing to count the model's purpose.
The base model is not broken. It was trained on MANE and is doing exactly what it was trained to do. The meta layer adds the context that lets it see past the canonical reference.
Gene navigation¶
The Go to gene box jumps to another gene, keeping the current base model. Genes without a prebuilt feature cache return a 404 naming the command to build one — see Getting started.