Novel Site Explorer (/novel/{gene})¶
A different question from the Genome View. That page asks "did the model get this gene's known sites right?" This one asks the discovery question:
For this gene, what are the top candidate sites that appear in no annotation at all?
That is a ranked list, and per-gene precision@k is literally the metric M3 is evaluated on — so the page has the same shape as the evaluation.
What a row shows¶
| Column | Meaning |
|---|---|
| Rank / M3 score | the meta model's ranking of this candidate |
| Base score, base rank | what the base model gave the same position |
| GT/AG | the splice dinucleotide, strand-aware |
| Evidence badges | ENCODE long-read support, held-out disease anchors |
The GT/AG column is a live correctness check, not decoration. Every row landing on the canonical dinucleotide for its strand and site type means the coordinate handling is right. It caught a real bug during development: offsets reasoned from first principles scored 0.00, and the convention had to be recovered empirically.
Evidence is deliberately independent¶
Badges come from ENCODE long-read transcripts (D1) and held-out disease anchors (D2). Neither is what M3 trained on.
SpliceVault is deliberately not shown
SpliceVault (positives_pooled) is M3's training pool. Displaying it as supporting evidence
would be circular — it would say "the model found what it was trained on."
The servable universe¶
Serving is restricted to the 4,956 genes on held-out chromosomes ⅓/5/7/9 — the ones M3 never trained on. 165 of them carry a novel disease anchor.
This is a feature. You cannot accidentally demo a training gene. Familiar genes like BRCA1, TP53 and UNC13A are on training chromosomes and are not available here; the page loads and shows an inline message naming the universe.
A good first gene: DHX29¶
| M3 rank | Type | M3 prob | Base prob | Base rank | GT/AG | Long-read |
|---|---|---|---|---|---|---|
| 1 | acceptor | 1.0000 | 0.0000 | 284 | AG | ✓ 8 biosamples |
| 2 | donor | 1.0000 | 0.0008 | 13 | GT | — |
| 3 | donor | 0.9999 | 0.0395 | 2 | GT | ✓ 1 |
The base model buries a real, long-read-confirmed, SF3B1-mutant cryptic site at rank 284 with probability ≈ 0. M3 ranks it first.
ATP6V1A is a good second (anchor at M3 rank 5 vs base rank 132). TPR works but is a weaker
illustration — base already ranks its anchors 3rd and 4th.
How to read a low base score¶
A near-zero base score is the point, not a failure. These are cryptic sites: low base score is what makes them cryptic. The gain here is re-ranking, not new evidence.
Equally, a top-ranked candidate is a candidate. Across the held-out set M3 recovers about 79% of disease cryptics in the top 20 per gene versus 51% for base, so expect roughly 1–2 of any top-5 to be long-read confirmed. It is a triage tool, not a truth oracle.
Status¶
M3 is registered status: research, not promoted. It is deliberately kept out of the genome-view
overlay dropdown and the metrics dashboard: it is scored by per-gene precision@k on independent
truth, not by the PR-AUC those surfaces expect, and its 7-channel input would fail against the
9-channel cache they use.
Full results: M3 — Novel Splice Sites.