Degradation-aware coverage evaluation: 2026-07-19

Follow-up: the intact/degraded kernel evaluated here has been superseded by the validated three-component competing-risk model, which separates length-invariant technical truncation from degradation.

Model

--coverage-model degradation --seq-tech ont-drna models the global sample-level process that generates 3'-anchored fragments before interpreting endpoint patterns as transcript evidence. For transcript length L and sequenced fraction u, the degraded component is a truncated exponential survival distribution with rate hazard_per_kb * L. An intact point component accounts for alignments in the first 5' gap bin.

The hazard is selected on a small regularized grid and the intact fraction is estimated continuously by MAP-EM using only single-candidate, 3'-anchored reads. Every read is scored with parameters trained without its fold. Candidate endpoint likelihood is divided by this global nuisance likelihood and renormalized.

The unified auto mode learns three levels of weighting. A sample-level null sets degradation correction to zero when a minimum-hazard technical process fits as well as degradation. Otherwise, each read is weighted by its posterior probability of degradation. Finally, normalized mixture variance 4 * intact_fraction * (1 - intact_fraction) controls whether both components are represented well enough to identify the correction. The same learned quantity attenuates endpoint reliability before the existing agreement gate and Bayes-factor cap.

This is deliberately an ONT direct-RNA kernel. The command requires an explicit --seq-tech ont-drna, including for BAM input, so it cannot silently apply the wrong directional model to cDNA or PacBio data.

Complete degradation trajectories

All 15 direct-RNA samples with public FASTQ in ENA PRJEB53210 were evaluated as fixed 50,000-read prefixes. Parameters were frozen before this expansion.

Trajectory	RIN sequence	Learned hazard sequence (/kb)	RIN/hazard Pearson
TS10	9.8, 9.6, 9.3, 8.4, 7.7	0.20, 0.20, 0.20, 0.36, 0.40	-0.969
TS11	9.7, 9.6, 9.3, 8.9, 8.8, 8.7	0.10, 0.10, 0.40, 0.70, 0.40, 0.40	-0.767
TS12	9.9, 8.2, 7.3, 7.2	0.20, 0.40, 0.70, 0.58	-0.948
pooled	15 samples	0.356 mean	-0.764

The hazard tracks lower RIN strongly in every trajectory. It is not perfectly monotone—most visibly around TS11 RIN 8.9—which is expected from 50k-read subsets and from RIN measuring total RNA rather than only sequenceable poly(A) molecules. The fitted intact fraction is less informative: it is 0.5 for 14/15 samples and 0.25 for TS11 RIN 8.9. The hazard, rather than mixture weight, is currently the identifiable degradation diagnostic.

High-to-low RIN robustness

Trajectory	Model	Pearson	Jensen-Shannon	L1 distance
TS10 9.8→7.7	none	0.9662	0.1415	0.6463
	adaptive	0.9542	0.1594	0.6825
	degradation	0.9605	0.1585	0.6780
	auto	0.9655	0.1515	0.6651
TS12 9.9→7.2	none	0.9468	0.1298	0.6313
	adaptive	0.9196	0.1538	0.6824
	degradation	0.9266	0.1526	0.6795
	auto	0.9479	0.1435	0.6541
TS12 9.9→7.3	none	0.9378	0.1283	0.6155
	adaptive	0.9118	0.1537	0.6698
	degradation	0.9200	0.1513	0.6616
	auto	0.9392	0.1423	0.6365

Automatic degradation-aware correction improves over adaptive on every trajectory and metric shown. On both TS12 comparisons its Pearson correlation slightly exceeds no coverage, while Jensen-Shannon and L1 remain between no coverage and adaptive. Since degradation genuinely changes observed abundance, this is a robustness test, not absolute transcript truth; nevertheless, the remaining gap is a reason to keep the model opt-in.

Frozen truth benchmarks

Dataset	Model	CCC	MARD	RMSE
large synthetic dRNA	none	0.956952	0.169178	89.4590
	adaptive	0.996935	0.143259	24.0550
	degradation	0.991278	0.143207	40.3891
	auto	0.996935	0.143259	24.0550
SIRV E0 dRNA 50k	none	0.639192	0.166302	0.529723
	adaptive	0.671348	0.154340	0.493311
	degradation	0.641909	0.165376	0.526606
	auto	0.671323	0.154345	0.493338
source-labelled dRNA	none	0.981515	0.0001168	0.012606
	degradation	0.982339	0.0001127	0.012317
	adaptive	0.983387	0.0001062	0.011939
	auto	0.983374	0.0001063	0.011944

The final automatic model retains adaptive's SIRV and large-synthetic accuracy: both select the minimum hazard and correction weight zero. On the nearly full-length source-labelled set, auto is numerically indistinguishable from adaptive because intact fraction is 0.998 and normalized mixture variance is near zero. Thus the degradation robustness no longer costs accuracy on any of the frozen truth benchmarks.

Across the 15 degradation samples, coverage construction averages 0.157 s and EM averages 1.235 s. This is about 0.12 s more coverage work than adaptive on the earlier six-sample panel. The large 1.36-million-read synthetic run spends 4.49 s on coverage and 16.95 s in EM.

Decision

• Prefer experimental auto over asking users to choose adaptive versus degradation. It preserves adaptive behavior when degradation is unsupported and continuously increases degradation correction when it is identifiable.
• Keep the overall default at none until the full multi-sample LongBench truth evaluation is complete; automatic model selection solves the interface and major accuracy tradeoff, but the external evidence base is still modest.
• The next model iteration should replace the coarse intact-fraction grid with a continuous or beta-binomial estimate and allow a third component for technical sequencing termination. Those components need sample-level outer validation before changing defaults.
• Technology separation is warranted. ONT cDNA and PacBio should receive their own observation kernels rather than reuse this 3'-anchored survival process.

Artifacts are under /scratch1/rob/long-read-ecosystem/oarfish-evaluation-data/degradation and /scratch1/rob/long-read-ecosystem/oarfish-evaluation-data/degradation-model-eval.