Skip to content

Meta-Layer Results & Findings

A single, reconciled view of how the meta-layer models perform — what each one is measured against, its current headline numbers, and the finding it establishes. Each model has its own page; this index is the cross-model summary and the guide to reading the numbers correctly.

How this relates to the workflow series

The Meta-Layer MLOps Workflow tells you how to produce and evaluate these models. This series tells you what came out — the evaluated results and the conclusions drawn from them. Where a number depends on a protocol, the model page links back to the relevant workflow stage.


Cross-model summary

Current, promoted models. All splice tasks are severely class-imbalanced, so the headline metric is PR-AUC (and, for the discovery tasks, precision@k / recall@k per gene), never accuracy alone. "vs base" is the improvement over the OpenSpliceAI base model on the same evaluation.

Model Task Evaluation Headline (current) vs base
M1-P Canonical, position-level (XGBoost) SpliceAI holdout PR-AUC donor 0.9986 / acceptor 0.9972 FN −62%, FP −68%
M1-S Canonical, sequence CNN SpliceAI holdout (MANE) macro PR-AUC 0.9998 0.9986 → 0.9998
M2-S Alternative sites (Ensembl ∖ MANE) Eval-Ensembl-Alt alt-site PR-AUC 0.990, recall ~90% 0.911 → 0.990, FN −88%
M3-S (v1) Novel sites (junction-supported) anti-circular D1 / D2 D1 P@5 0.335, D2 R@20 0.79 beats base (0.277 / 0.51)
M3-R Candidate refiner (rerank) anti-circular D1 / D2 ties base (honest negative) within-gene flat
M4 Perturbation-induced in progress

The one-line story: the canonical task (M1) is essentially solved; the alternative-site task (M2) is where multimodal refinement pays off most dramatically; and the novel-site frontier (M3) is where a learned recognizer (M3-v1) helps but genome-averaged multimodal evidence, by itself, does not (M3-R).


How to read these numbers

A few conventions are applied consistently across every page, so the results can be compared honestly.

Metric choice. Splice sites are ~0.1–1% of positions. We report PR-AUC (ranking quality independent of threshold) and, for discovery, precision@k / recall@k per gene. Accuracy is quoted only where it adds context.

Operating points are labeled. Raw false-negative / false-positive counts depend entirely on the decision threshold, and at the naïve argmax/0.5 threshold they are misleading under this imbalance. So when counts appear, they are reported at the F1-optimal (or a stated matched-recall) operating point, and the operating point is named. The same run at argmax will show very different FP counts — that is a threshold artifact, not a ranking change.

"Anti-circular" evaluation (M3). For novel-site discovery, evaluating against the annotation the model was trained on is circular. M3 results are reported on independent truth sets — ENCODE long-read novel junctions (D1) and held-out disease anchors (D2) — with the training/annotated sites subtracted out. See M3 novel-site results.

Model naming. M{task}-{S|P}: task 1–4 (canonical / alternative / novel / perturbation), -S = sequence model, -P = position-level. Full convention in naming_convention.md.


Model status

Model Version Status
M1-P full-genome Reference baseline
M1-S m1s_v4_cleanannot Promoted (canonical)
M2-S m2s_v4_cleanannot Promoted (alternative)
M3-S m3_v1 Best novel-site ranker (research)
M3-S v1.1 m3s_v1_1_confirmed Tier 1 confirmed-only retrain — marginally worse than v1 (label noise was not the ceiling)
M3-R m3r_candidate_refiner Honest negative — not promoted; global-AUC triage repurposable
M4 In progress (perturbation-paired labels needed)

Superseded versions (M1-S v1/v2 blends, M2-S v1/v2) are summarized as version history on each model page — the design trail is preserved without competing with the current numbers.


Pages

Sources. Curated write-ups in examples/meta_layer/results/ and examples/meta_layer/docs/M3/; machine-readable metrics under output/meta_layer/<model>/.