Skip to content

Evaluation checkpoint: EM acceleration and endpoint model

Superseded endpoint result: the 2026-07-19 evaluation found that endpoint probabilities were trained but not enabled in the EM. Endpoint conclusions in this checkpoint are invalid; see coverage-evaluation-2026-07-19.md.

Date: 2026-07-18. This is an implementation checkpoint, not a promotion report. Commands used the release build from the working tree and fixed input alignments or identical raw-read mapping options between model variants.

Bundled alignment smoke test

Input: eval/parity/small_txp.bam (4,198 aligned reads; 9,579 retained alignments), --filter-group no-filters, four threads, coverage disabled.

EM method M-step evaluations Converged at 0.001
ordinary EM 1,000 no
SQUAREM 223 yes
DAAREM 406 yes

The three output totals agree to floating-point precision (4,198). Their truth-comparison metrics are indistinguishable at the displayed precision. The available truth file describes a much deeper library than this small BAM, so it is useful for numerical parity after library-size scaling but is not treated as an accuracy benchmark.

Local direct-RNA checkpoint

Input: drna_test/hek_100k.fastq, GENCODE v47 transcriptome, ONT direct-RNA preset. Ordinary and repaired-logistic runs completed on 95,324 mapped reads. The full endpoint run was repeatedly terminated by the execution environment during raw-read mapping, before endpoint-model training. A 2,000-read subset completed and trained on 781 unambiguous reads. This is recorded as an unresolved resource/execution failure; it is not evidence for or against the model's accuracy.

Public SIRV checkpoint

Public ONT direct-RNA SIRV mixtures were downloaded to /scratch1/rob/long-read-ecosystem/oarfish-evaluation-data/public-sirv:

Mixture Run FASTQ MD5 Compressed bytes
E2 SRR6058583 1146f3fbd028d1b6502cb4dd38b8453e 222,870,088
E0 SRR6058584 ff98d2f899dcf52774b55310fc70b268 120,341,594

The matching Lexogen reference FASTA MD5 is 69cf02d2cd69ff118967e291c00df21a. Transcriptome-mode runs against its 114 reference records completed for all three coverage modes. On E0, using the known equal isoform mixture aggregated to the seven SIRV loci, ordinary EM and endpoint were identical (CCC 0.948587, Pearson 0.978646, RMSE 0.940893); logistic changed only in the sixth decimal place (CCC 0.948589, RMSE 0.940877).

This is only a locus-level sanity check: the distributed FASTA contains SIRV loci plus ERCC records, while isoforms are represented in the matching GTF. Genome-projection runs were terminated during mapping in this environment. Therefore these results do not satisfy the transcript-level promotion gate.

Decision

  • Keep --em-accel none and --coverage-model none as defaults.
  • Accept SQUAREM and DAAREM as opt-in inference features; both reached the same numerical solution with fewer map evaluations in the bundled smoke test.
  • Retain endpoint as an experimental opt-in model. Do not implement or promote endpoint mixtures, predictive coverage, or segment models until a held-out, transcript-truth evaluation and the local full-data failure are resolved.