Datasetpublic

depinwang/lsec-sicilian-lonp2-pilot-results-v1

lsec-sicilian-lonp2-pilot-results-v1 Pilot results (5 samples) from re-detecting splice junctions with SICILIAN (Dehghannasiri, Olivieri, Salzman -- Genome Biology 2021) on raw FASTQ, to test whether the LONP2 skipped-exon (SE) event's null result (3x via STARsolo -> JAseC/SHIBA) is a methods artifact. See experiment lsec-sicilian-lonp2 for the full design, red-team review, and debugging history. The LONP2 SE event (hg38) 5' constitutive exon:… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-sicilian-lonp2-pilot-results-v1.

sourceHugging Faceupdated 20d agoView on Hugging Face

0likes42downloads

Dataset Card

lsec-sicilian-lonp2-pilot-results-v1

Pilot results (5 samples) from re-detecting splice junctions with SICILIAN (Dehghannasiri, Olivieri, Salzman -- Genome Biology 2021) on raw FASTQ, to test whether the LONP2 skipped-exon (SE) event's null result (3x via STARsolo -> JAseC/SHIBA) is a methods artifact. See experiment lsec-sicilian-lonp2 for the full design, red-team review, and debugging history.

The LONP2 SE event (hg38)

—5' constitutive exon: chr16:48252131-48252365
—Alternative (skipped) exon: chr16:48256610-48256741
—3' constitutive exon: chr16:48258618-48258740
—Exclusion junction (skips the alt exon): chr16:48252365-48258618 (SICILIAN exon-boundary convention) / chr16:48252366-48258617 (STAR SJ.out.tab intron-boundary convention)

Samples

5 samples: fibrosis1/S38 (SRR28058446) and fibrosis4/S33 (SRR28058449) -- the only 2 samples with any current-pipeline read support at the LONP2 SE junction -- plus 3 Healthy controls (SRR28058450/451/452).

Columns

Column	Description
`srr_id`	SRA run accession
`sample_label`	Short sample name used throughout this workspace
`group`	F2-3 (fibrosis) or Healthy
`raw_star_exclusion_junction_reads`	Unique-read count for the exclusion junction directly from STAR's `SJ.out.tab`, BEFORE any UMI deduplication or LSEC-cell restriction -- a cheap, SICILIAN-stats-independent check
`sicilian_exclusion_junction_n_barcodes`	Number of distinct cell barcodes with a SICILIAN-called (GLM-confidence-scored) exclusion-junction event, from `sicilian_called_splice_juncs.tsv`
`sicilian_exclusion_junction_sum_numreads`	Sum of the `numReads` field across those barcodes
`lonp2_sj_hits_raw_tsv`	Full raw content of `lonp2_sj_hits.tsv` (every STAR `SJ.out.tab` row in the LONP2 genomic window, not just the exclusion junction)
`lonp2_sicilian_hits_full_tsv`	Full raw content of `lonp2_sicilian_hits.tsv` (every row of `sicilian_called_splice_juncs.tsv` matching gene name LONP2, i.e. all LONP2 junctions SICILIAN called for this sample, not just the exclusion junction)

Key finding

The raw STAR SJ.out.tab counts for the exclusion junction show a clean group separation (F2-3: 45, 42 vs Healthy: 28, 23, 15 -- no overlap). But SICILIAN's own final, cell/UMI-aware, statistically-scored output (after its GLM confidence model) calls the exclusion junction in only ONE sample (S38: 2 barcodes, 4 total reads) -- zero in S33 and all 3 Healthy samples. Most of the raw-read-level signal did not survive SICILIAN's own candidate-junction detection and statistical filtering. This largely REPRODUCES the original STARsolo->JAseC/SHIBA pipeline's near-null result rather than revealing a methods-artifact-masked signal, though the single positive sample is directionally consistent with the original hypothesis (it's an F2-3 sample).

Not included in this dataset (too large, kept on-cluster for reproducibility)

Per-sample class_input.tsv (~14-16GB), GLM_output.txt (~200-450MB), and the full sicilian_called_splice_juncs.tsv (330MB-3.1GB) remain on turso at /wrk-kappa/users/depinwan/tries/2026-05-05-rerun_Liver_dataset_from_Nature/lsec_sicilian_lonp2/pilot/<SRR>/.