Competing-risk degradation evaluation: 2026-07-20
This stage tested two frozen alternatives to the original ONT direct-RNA degradation kernel: internally learned scalar correction weights and an observation model that separates degradation from technical truncation.
Rejected: alignment-predictive scalar calibration
The original intact/degraded model was evaluated at correction weights
0, 0.1, 0.25, 0.5, 0.75, 1. Cross-fitted ambiguous reads were scored by the
agreement of corrected endpoint probabilities with their independent alignment
score probabilities. A conservative prior favored zero and likelihood evidence
was capped at 50 effective reads.
This successfully detected LongBench overcorrection. Learned weights were 0.03--0.07 in the four active 50,000-read libraries, and HCC827 at 250,000 reads improved from CCC 0.558229 to 0.562535, essentially equal to adaptive (0.562512). It failed the degradation-trajectory gate, however:
| Trajectory | Old auto Pearson | Predictive-weight Pearson |
|---|---|---|
| TS10 9.8→7.7 | 0.96549 | 0.95439 |
| TS12 9.9→7.2 | 0.94790 | 0.91991 |
| TS12 9.9→7.3 | 0.93919 | 0.91205 |
The selector assigned the degradation series the same small weights as LongBench, collapsing robustness almost to adaptive. Alignment agreement is therefore not an adequate single-sample proxy for quantification benefit. The selector was removed rather than retained as unused runtime overhead.
Accepted: three-component competing-risk model
The accepted model represents a 3'-anchored read as one of:
- intact, concentrated in the full-length 5' endpoint bin;
- technical truncation, a length-invariant Beta(2,1) retained fraction; or
- degradation, a truncated exponential whose rate scales with physical transcript length.
Mixture weights are estimated continuously by cross-fitted MAP-EM with a weak Dirichlet(5,2,2) prior. Hazard is selected on the frozen regularized grid. The correction divides only the degradation likelihood ratio relative to the intact/technical mixture, rather than removing the entire observed endpoint profile. Its strength remains learned from degradation-versus-null evidence, the degraded mixture fraction, and each read's degradation posterior. Existing adaptive reliability and Bayes-factor safeguards remain in force.
Across LongBench dRNA, the model estimates 7.6--20.1% degraded molecules and 20.0--26.5% technical truncation at 50,000 reads. These components are separated by their different transcript-length dependence rather than endpoint shape alone.
Results
Degradation robustness
| Trajectory | None Pearson | Old auto Pearson | Competing-risk Pearson | Competing-risk JSD | Competing-risk L1 |
|---|---|---|---|---|---|
| TS10 9.8→7.7 | 0.96623 | 0.96549 | 0.96500 | 0.15519 | 0.66866 |
| TS12 9.9→7.2 | 0.94682 | 0.94790 | 0.94876 | 0.14290 | 0.64831 |
| TS12 9.9→7.3 | 0.93781 | 0.93919 | 0.94013 | 0.14257 | 0.63269 |
The new model preserves the original auto robustness and slightly improves both TS12 comparisons.
LongBench matched-Illumina comparator
At 50,000 reads, competing-risk auto is within 0.00027 CCC of adaptive or
better in all eight cell lines, while every cell line remains above none.
It improves over the old auto in each library where old degradation correction
was active. At 250,000 HCC827 reads:
| Model | CCC | Pearson | RMSE |
|---|---|---|---|
| adaptive | 0.562512 | 0.739364 | 6869.94 |
| old auto | 0.558229 | 0.735039 | 6919.07 |
| competing-risk auto | 0.562581 | 0.739437 | 6869.19 |
Frozen truth
| Dataset | CCC | MARD | RMSE |
|---|---|---|---|
| large synthetic dRNA | 0.996926 | 0.143267 | 24.0920 |
| SIRV E0 dRNA | 0.671364 | 0.154332 | 0.493293 |
| source-labelled dRNA | 0.983387 | 0.0001062 | 0.011939 |
These are effectively equal to the frozen adaptive optima. The model assigns only 0.69% degraded fraction to SIRV and 0.22% to the nearly full-length source-labelled set, making correction negligible without a manually selected mode.
Runtime and decision
Caching component probabilities during mixture EM reduced coverage construction on the 1.36-million-read synthetic benchmark from 11.3 s during development to 2.66 s; EM took 17.45 s. This is also below the previous two-component auto measurement of 4.49 s. The cached and uncached implementations produce the same fitted parameters and estimates.
Retain the three-component constant-hazard model. Bounded beta and piecewise challengers were subsequently tested; neither displaced the incumbent. See the challenger study. Gene-random effects and joint abundance/degradation inference remain unjustified.
Artifacts are under oarfish-evaluation-data/degradation/three-component* and
oarfish-evaluation-data/longbench/three-component*.