Rank-preserving adaptive inference evaluation (2026-07-22)
Motivation
The automatic coverage, censoring, and dominance models improve count-scale metrics, but their largest remaining weakness was low-abundance rank order. A coverage-free preliminary EM estimate often preserved that order. We therefore tested an abundance-dependent blend between the preliminary and corrected estimates.
The retained blend keeps at least 80% of the corrected estimate. Its correction fraction rises smoothly to 100% with preliminary abundance, with a midpoint of 300 counts per million reads. This primarily protects low-abundance ranks while leaving high-abundance coverage corrections nearly unchanged.
Selection experiments
- A global blend improved Spearman on all 28 primary cases (mean +0.00822 versus the previous automatic model), but erased much of the coverage-model benefit on independent ONT and PacBio simulations. It was rejected as a default.
- Transcript-level gating by unique-read support also failed those independent simulations and was rejected.
- The censoring scale learned from unique reads separated the cases. Independent ONT/PacBio simulations fit the lower clamp of 25 nt; the main datasets fit 39--92 nt. SIRV mixtures fit 25--36 nt but contain only 69 transcripts.
- The retained automatic rule activates rank preservation only for transcriptomes with at least 1,000 transcripts and a learned censoring scale above 37 nt. Otherwise it abstains and uses the corrected EM estimate unchanged.
The thresholds are deliberately based only on sample-internal diagnostics; no
truth labels enter inference. --rank-blend fixed remains available for explicit
experiments, and --rank-blend none disables the additional warm-up.
Accuracy
On the 28-case primary panel, automatic selection activates for every case and matches the validated fixed blend:
| Comparison | Pearson | Spearman | CCC | RMSE | MARD |
|---|---|---|---|---|---|
| vs previous auto, mean delta | +0.00021 | +0.00822 | -0.00052 | +15.36 | -0.00090 |
| wins vs previous auto | 16/28 | 28/28 | 4/28 | 3/28 | 20/28 |
| vs no coverage, mean delta | +0.02314 | +0.00471 | +0.02296 | -155.00 | +0.00037 |
| wins vs no coverage | 27/28 | 27/28 | 27/28 | 25/28 | 9/28 |
Positive is better for correlations; negative is better for RMSE and MARD. The remaining Spearman loss versus no coverage is H69 PacBio (-0.00186), while the model retains its count-scale benefit there.
On outer validation the rule abstains in all five cases. It therefore retains the previous model's strong results instead of the rejected global blend:
| Sample | Previous/automatic Spearman | Global blend Spearman |
|---|---|---|
| independent ONT simulation | 0.7679 | 0.7048 |
| independent PacBio simulation | 0.7642 | 0.7099 |
| SIRV E2 dRNA | 0.8749 | 0.8692 |
Cost
When active, the preliminary coverage-free EM adds a mean 0.079 seconds and about 1.0 MiB peak RSS versus the previous automatic model on the primary panel. When the learned rule abstains, only a linear scan over unique alignments is added; no preliminary EM or blend allocation is performed.
Artifacts are in oarfish-evaluation-data/rank-blend-mid-300-full-20260722,
rank-blend-auto-outer-20260722, and rank-blend-auto-boundary-20260722.