datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
alternative-splicing-glm-tang
GLM-Tang Task 4: Tissue-specific Alternative Splicing
This dataset contains the cassette-exon prediction task used in Task 4 of
Tang et al. Given sequence around the splice acceptor and donor, predict
tissue-specific exon inclusion across 56 tissues.
Loading
from datasets import load_dataset
dataset = load_dataset("Taykhoom/alternative-splicing-glm-tang")
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]
Dataset size… See the full description on the dataset page: https://huggingface.co/datasets/Taykhoom/alternative-splicing-glm-tang.image-splicing-deepfake-mix-newlsec-fcn2-splicing-viz-valerie-plot-v1
lsec-fcn2-splicing-viz-valerie-plot-v1
VALERIE v2.1.2 PlotPSI output for the FCN2 SE event across all 17 samples / 4 groups (1666 split LSEC cells). method=kw (Kruskal-Wallis; wilcox/t.test crash on >2 factor levels in v2.1.2). Per-group split cells with >=2 region reads (coverage proxy): F0=257,F2-3=470,F4=62,Healthy=726.
Dataset Info
Rows: 2
Columns: 6
Columns
Column
Type
Description
image
Image(mode=None, decode=True)
PNG plot from… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-fcn2-splicing-viz-valerie-plot-v1.lsec-fcn2-splicing-viz-percell-counts-v1
lsec-fcn2-splicing-viz-percell-counts-v1
Per-cell read-level junction counts for the FCN2 SE event, from lsec_bams_dedup (UMI-dedup) split by CB in cell_barcode.tsv.gz. E/I = read counts of the exclusion / inclusion junctions; I = max(inclA, inclB); PSI = I/(I+E). Read-level, NOT UMI-collapsed. Complete 17-sample set from the scale-up re-split (3 canary + 14 added samples), the same rows the full 4-group bampheno consumed.
Dataset Info
Rows: 989
Columns: 8… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-fcn2-splicing-viz-percell-counts-v1.jinyang-gse138866-splicing-subtypes-results-v1
jinyang-gse138866-splicing-subtypes-results-v1
Canary run (N=110 real GSE138866 samples) for AS-based consensus-clustering molecular
subtype discovery. Outcome: GATE_FAILURE (null result). Only 2/10,019 high-confidence
AS events clear BH-FDR<0.05 (need >=20 to proceed to clustering) -- a pre-specified valid
stopping point ("insufficient power to detect stable AS-based subtypes"), not a pipeline bug.
Downstream clustering/KM/permutation-null artifacts were not produced, per… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/jinyang-gse138866-splicing-subtypes-results-v1.lsec-fcn2-splicing-viz-pseudobulk-psi-v1
lsec-fcn2-splicing-viz-pseudobulk-psi-v1
JAseC-faithful pseudobulk PSI replication for the FCN2 SE event, all 17 samples. E/I computed with JAseC's exact counting rules (bamJuncCount_10x: mapped, N-containing CIGAR, CB+UB tags, NH==1, no secondary/supplementary filter, UMI collapse per (junction,strand,cell), antisense merge at same jc_pos). match column compares E_ref/I_ref from the JAseC reference run: 17/17 exact.
Dataset Info
Rows: 17
Columns: 17… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-fcn2-splicing-viz-pseudobulk-psi-v1.image-splicing-deepfake-mixlsec-biomarkers-splicing-viz-mt1g-plot-v1
lsec-biomarkers-splicing-viz-mt1g-plot-v1
VALERIE v2.1.2 PlotPSI output for the MT1G event (851 split LSEC cells, cell.types=Healthy,F0,F2-3,F4, method=kw). Per-group split cells with >=2 region reads (coverage proxy): Healthy=410,F0=211,F2-3=219,F4=5.
Dataset Info
Rows: 2
Columns: 6
Columns
Column
Type
Description
image
Image(mode=None, decode=True)
PNG plot from PlotPSI: per-cell coverage-ratio PSI heatmap, mean PSI +/- bootstrap CI… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-biomarkers-splicing-viz-mt1g-plot-v1.lsec-lonp2-splicing-viz-valerie-plot-v1
lsec-lonp2-splicing-viz-valerie-plot-v1
VALERIE v2.1.2 PlotPSI output for the LONP2 SE event, HC vs F2-3 (315 split LSEC cells). method=wilcox (2 groups). Per-group split cells with >=2 region reads (coverage proxy): Healthy=87,F2-3=45. KNOWN-NULL LOCUS — see event description. This is a negative-control companion to lsec-fcn2-splicing-viz-valerie-plot-v1, not a positive finding: expect near-uniform PSI=1.00 and a flat/non-significant p-track.
Dataset Info
Rows:… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-lonp2-splicing-viz-valerie-plot-v1.lsec-biomarkers-splicing-viz-stab2-plot-v1
lsec-biomarkers-splicing-viz-stab2-plot-v1
VALERIE v2.1.2 PlotPSI output for the STAB2 event (1929 split LSEC cells, cell.types=Healthy,F0,F2-3,F4, method=kw). Per-group split cells with >=2 region reads (coverage proxy): Healthy=779,F0=426,F2-3=478,F4=166.
Dataset Info
Rows: 2
Columns: 6
Columns
Column
Type
Description
image
Image(mode=None, decode=True)
PNG plot from PlotPSI: per-cell coverage-ratio PSI heatmap, mean PSI +/- bootstrap… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-biomarkers-splicing-viz-stab2-plot-v1.lsec-biomarkers-splicing-viz-stab2-percell-counts-v1
lsec-biomarkers-splicing-viz-stab2-percell-counts-v1
Per-cell region-read coverage table for the STAB2 event, all 17 samples. Per-cell coverage-proxy table (region_reads only -- RI PSI is computed by VALERIE internally from per-base read-coverage ratios over the intron span, not from junction counts we compute ourselves; see EXPERIMENT_README.md section 4).
Dataset Info
Rows: 1929
Columns: 4
Columns
Column
Type
Description
cb… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-biomarkers-splicing-viz-stab2-percell-counts-v1.lsec-lonp2-splicing-viz-percell-counts-v1
lsec-lonp2-splicing-viz-percell-counts-v1
Per-cell read-level junction counts for the LONP2 SE event, HC vs F2-3 (10 samples), from lsec_bams_dedup (UMI-dedup) split by CB in cell_barcode.tsv.gz. E/I = read counts of the exclusion / inclusion junctions; I = max(inclA, inclB); PSI = I/(I+E). Read-level, NOT UMI-collapsed. Known-null locus — expect PSI=1.00 for nearly all cells in both groups.
Dataset Info
Rows: 38
Columns: 8
Columns
Column… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-lonp2-splicing-viz-percell-counts-v1.lsec-biomarkers-splicing-viz-mt1g-percell-counts-v1
lsec-biomarkers-splicing-viz-mt1g-percell-counts-v1
Per-cell region-read coverage table for the MT1G event, all 17 samples. Per-cell coverage-proxy table (region_reads only -- A3SS PSI is computed by VALERIE internally from per-base read-coverage ratios, not from junction counts we compute ourselves; see EXPERIMENT_README.md section 4).
Dataset Info
Rows: 851
Columns: 4
Columns
Column
Type
Description
cb
Value('string')
cell barcode… See the full description on the dataset page: https://huggingface.co/datasets/depinwang/lsec-biomarkers-splicing-viz-mt1g-percell-counts-v1.lsec-fcn2-splicing-viz-valerie-plot-v1splicing_epistasis_okgp_cons_silent_control
CONSTRAINED TCGA-silent control (Panel F)
84,848 cis-double SNV pairs across 261 randomly sampled genes: LOEUF<0.7, ZERO TCGA pairs in any tier. Null reference.
Built with the standard paper §2 ¶1-5 methodology:
≤100 nt intra-gene window
≥3 1KG carriers per constituent
TopLD r² ≥ 0.2 dropped
Greedy 1:1 matched arms keyed on gene + log-rarity bin + distance ±2 nt
seed=42
Companion panels:
splicing_epistasis_okgp_cons_driver_tcgarec — Panel D… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_cons_silent_control.splicing_epistasis_okgp_tcga_driver_recurrent
Splicing epistasis — Driver+TCGA-recurrent panel (Panel A)
12,714 cis-double SNV pairs (6,357 observed + 6,357 unobserved-matched) in 1000 Genomes
across 56 genes that satisfy:
listed in OncoKB cancerGeneList.tsv (any inclusion criterion), AND
carry ≥2 TCGA-clean double-mutation pairs each with ≥5 carriers
(per tcga_de_novo_pairs_clean.csv from the original paper-2 code dump)
Built as the driver arm of a three-panel strict test that disentangles cancer-driver
status from… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_tcga_driver_recurrent.splicing_epistasis_okgp_noncancer1000
Splicing epistasis — non-cancer 1000-gene control set
167,322 cis-double SNV pairs (83,661 observed + 83,661 unobserved-matched) in 1000 Genomes,
enumerated within a 1000-gene random sample of protein-coding genes excluding all OncoKB entries.
Built as a control comparator for the OncoKB-987-driver dataset to test whether the splicing-epistasis
depletion signal (Lynn et al., Section 2 ¶1-5) is panel-specific or generalizes to non-cancer genes.
Same methodology as paper §2: ≤100… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_noncancer1000.athaliana_rnaseq_splicing
Arabidopsis thaliana RNA-seq Splicing Dataset
Dataset Description
This dataset contains splice junction information from Arabidopsis thaliana RNA-seq data across multiple tissues (shoot, root, leaf, flower).
Data Processing
RNA-seq data mapped to the reference genome using STAR
Splice junctions filtered based on supporting read counts
Junctions with ±2bp shifts merged using bedtools
Data split by chromosomes: chr 1-3 for training, chr 4 for validation, and chr… See the full description on the dataset page: https://huggingface.co/datasets/suzuki-2001/athaliana_rnaseq_splicing.splicing_epistasis_okgp_oncokb987
Splicing epistasis — OncoKB-driver-extended 1000 Genomes matched-arms set
262,040 cis-double SNV pairs in 1000 Genomes, enumerated within the 987 OncoKB-annotated cancer driver genes, split into matched observed / unobserved arms for the splicing-epistasis depletion test (Lynn et al., Genome Medicine, in preparation).
Why this dataset
The paper's headline depletion test (Section 2 ¶1-5) runs on a 138-gene cancer panel and 28,642 matched-arms pairs. This dataset… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_oncokb987.splicing_epistasis_okgp_cons_driver_tcgarec
CONSTRAINED OncoKB drivers with TCGA-recurrent doublets (Panel D)
10,864 cis-double SNV pairs across 41 genes: OncoKB cancer drivers, LOEUF<0.7, with ≥2 TCGA-clean pairs × ≥5 carriers each. Tests the RESCUE-enrichment hypothesis.
Built with the standard paper §2 ¶1-5 methodology:
≤100 nt intra-gene window
≥3 1KG carriers per constituent
TopLD r² ≥ 0.2 dropped
Greedy 1:1 matched arms keyed on gene + log-rarity bin + distance ±2 nt
seed=42
Companion panels:… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_cons_driver_tcgarec.splicing_epistasis_okgp_tcga_nondriver_recurrent
Splicing epistasis — Non-driver+TCGA-recurrent panel (Panel B)
94,984 cis-double SNV pairs (47,492 observed + 47,492 unobserved-matched) in 1000 Genomes
across 663 genes that satisfy:
NOT in OncoKB cancerGeneList.tsv, AND
carry ≥2 TCGA-clean double-mutation pairs each with ≥5 carriers
(per tcga_de_novo_pairs_clean.csv from the original paper-2 code dump)
Built as the non-driver arm of a three-panel strict test. These are genes that show
recurrent doublet events in TCGA cancers… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_tcga_nondriver_recurrent.splicing_epistasis_okgp_tcga_silent_control
Splicing epistasis — TCGA-silent control panel (Panel C)
143,408 cis-double SNV pairs (71,704 observed + 71,704 unobserved-matched) in 1000 Genomes
across 719 randomly sampled protein-coding genes that have zero TCGA double-mutation pairs
in tcga_de_novo_pairs_clean_relaxed.csv (the broad TCGA pair pool).
Built as the null arm of a three-panel strict test. These genes show no evidence of recurrent
doublet events in TCGA — the prediction is that they should show no 1KG depletion… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_tcga_silent_control.splicing_epistasis_okgp_test
Splicing epistasis — 1000 Genomes matched-arms test set
End-to-end reproducibility test set for the splicing-epistasis paper (Lynn et al., Genome Medicine, in preparation). 28,642 cis-double SNV pairs from a 138-gene cancer panel, matched in two arms (observed in 1000 Genomes vs. matched unobserved control).
Test purpose
This dataset exists to validate the engine-inference layer of the paper's pipeline. The 4-engine splice-prediction ensemble (Pangolin… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_test.splicing_epistasis_okgp_cons_nondriver_tcgarec
CONSTRAINED non-drivers with TCGA-recurrent doublets (Panel E)
41,522 cis-double SNV pairs across 220 genes: NOT in OncoKB, LOEUF<0.7, with ≥2 TCGA-clean pairs × ≥5 carriers each. Tests the SYNERGY-depletion hypothesis.
Built with the standard paper §2 ¶1-5 methodology:
≤100 nt intra-gene window
≥3 1KG carriers per constituent
TopLD r² ≥ 0.2 dropped
Greedy 1:1 matched arms keyed on gene + log-rarity bin + distance ±2 nt
seed=42
Companion panels:… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_cons_nondriver_tcgarec.Human-splicing-variantsscRNAseq-splicingsplicing_assaysplicing_epistasis_okgp_panelfree
splicing_epistasis_okgp_panelfree
Panel-free 1KG cis-pair universe for the splicing-epistasis paper §2A patch.
Filters
1000 Genomes phase-3, cis-pair within 100 nt
both constituents ≥ 3 carriers
d_ss_min ≤ 100 nt (splice-relevant subset)
TopLD-overlap excluded (haplotype-block confound)
Rows
124,888 pairs across 1,746 genes
arm split: {'unobserved': 109126, 'observed': 15762}
Columns
epistasis_id, arm, gene, chrom, mut1, mut2… See the full description on the dataset page: https://huggingface.co/datasets/nicolynnvila/splicing_epistasis_okgp_panelfree.