Taykhoom/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

Split	Rows	Evaluation-eligible rows
train	38,028	5,343
validation	1,088	884
test	11,840	1,621
Total	50,956	7,848

The validation chromosome split originally contains 11,955 chr1/7/9 exons. Tang's distributed validation CSV additionally applies:

nanstd(PSI / 100) > 0.1

which leaves the 1,088 rows published here.

Model input

The model receives two transcript-oriented 400 nt inputs:

acceptor_sequence = 300 nt upstream intron + 100 nt exon
donor_sequence    = 100 nt exon + 300 nt downstream intron

sequence is the serialized concatenation:

acceptor_sequence + donor_sequence

The two genomic arms overlap in 42,374 rows, are contiguous in 103, and have a gap in 8,479. Generic sequence models should consume the two explicit sequence columns or otherwise respect the serialization boundary at index 400.

Target representation

For each exon and tissue:

1. source PSI -1 becomes missing;
2. percentage PSI is divided by 100;
3. PSI is clipped to [1e-5, 1 - 1e-5];
4. observed tissue PSI values are logit transformed;
5. the target is:

target_mean_logit = mean(logit(observed tissue PSI))
label[tissue] = logit(PSI[tissue]) - target_mean_logit

label is a fixed-size 56-value vector. label_mask is a fixed-size boolean vector and is mandatory: 18,791 rows contain at least one missing tissue, totaling 321,222 masked target values.

Tang's released masked loss zeros missing entries and then averages across the entire batch-by-56 tensor:

tang_target = np.asarray(batch["aux_target_tang_delta_logit"])
observed = np.asarray(batch["label_mask"], dtype=bool)
per_entry_loss = tang_elementwise_loss(prediction, tang_target, batch)
loss = np.where(observed, per_entry_loss, 0).mean()

This preserves Tang's missingness-dependent loss scale. Dividing per_entry_loss[observed].sum() by observed.sum() is a valid observed-only-normalized alternative, but it is not the Tang-compatible reduction and should be reported as a different training protocol.

Absolute PSI logits can be reconstructed as:

absolute_logit = label + target_mean_logit
psi = sigmoid(absolute_logit)

Tang's preprocessing notebook used a different training target, logit(PSI) - logit(mean PSI). It is retained as aux_target_tang_delta_logit, and the processing-only delta_logit.h5 intermediate preserves the exact Tang layout. The primary label follows the pinned MTSplice evaluator, whose archived x array uses logit(PSI) - mean(logit PSI).

Tang's scalar logit(mean PSI) baseline is derivable for observed tissues as:

tang_logit_mean_psi = (
    label + target_mean_logit - aux_target_tang_delta_logit
)

The released model trains with the Tang auxiliary target and a masked KL-like loss. To reproduce the pinned evaluator, iterate over tissues and compute Spearman correlation between prediction and label[tissue] using only test rows for which both evaluation_included == True and label_mask[tissue] == True. Do not score a missing tissue merely because the exon is evaluation eligible. The paper figure text calls the metric Pearson, so users should state explicitly which protocol they follow.

The pinned MTSplice prediction artifact reproduces 56 per-tissue Spearman correlations on 1,621 eligible test rows: mean 0.25733, median 0.26145, minimum 0.11180, and maximum 0.40795.

Tissue order

The vector uses this complete zero-based order:

 0  Retina - Eye
 1  RPE/Choroid/Sclera - Eye
 2  Subcutaneous - Adipose
 3  Visceral (Omentum) - Adipose
 4  Adrenal Gland
 5  Aorta - Artery
 6  Coronary - Artery
 7  Tibial - Artery
 8  Bladder
 9  Amygdala - Brain
10  Anterior cingulate - Brain
11  Caudate nucleus - Brain
12  Cerebellar Hemisphere - Brain
13  Cerebellum - Brain
14  Cortex - Brain
15  Frontal Cortex - Brain
16  Hippocampus - Brain
17  Hypothalamus - Brain
18  Nucleus accumbens - Brain
19  Putamen - Brain
20  Spinal cord (C1) - Brain
21  Substantia nigra - Brain
22  Mammary Tissue - Breast
23  EBV-xform lymphocytes - Cells
24  Leukemia (CML) - Cells
25  Xform. fibroblasts - Cells
26  Ectocervix - Cervix
27  Endocervix - Cervix
28  Sigmoid - Colon
29  Transverse - Colon
30  Gastroesoph. Junc. - Esophagus
31  Mucosa - Esophagus
32  Muscularis - Esophagus
33  Fallopian Tube
34  Atrial Appendage - Heart
35  Left Ventricle - Heart
36  Cortex - Kidney
37  Liver
38  Lung
39  Minor Salivary Gland
40  Skeletal - Muscle
41  Tibial - Nerve
42  Ovary
43  Pancreas
44  Pituitary
45  Prostate
46  Not Sun Exposed - Skin
47  Sun Exposed (Lower leg) - Skin
48  Ileum - Small Intestine
49  Spleen
50  Stomach
51  Testis
52  Thyroid
53  Uterus
54  Vagina
55  Whole Blood

This card is sufficient to interpret the vectors. The processing contract is the machine-readable source for the same order, and verification requires this display to match it.

Columns

Column group	Description
split, example_id, source_index	Released partition, ASCOT exon ID, and source CSV index.
sequence, acceptor_sequence, donor_sequence, sequence_layout	Serialized and explicit two-arm model inputs.
arm_relationship, arm_overlap_or_gap_length	Whether the genomic arm windows overlap, are contiguous, or have a gap, plus the corresponding length (zero when contiguous).
label, label_mask	Fixed-size 56-value delta-logit target and observation mask.
target_mean_logit, target_observed_tissue_count	Evaluator mean-logit baseline and number of observed tissues; not input features.
aux_target_tang_delta_logit	Tang notebook target, retained for historical training reproduction; not an input feature.
evaluation_included	At least 10 observed tissues and maximum PSI deviation from the exon mean of at least 0.2.
chrom, start, end, strand	Zero-based, half-open hg38 exon coordinates and transcript strand.
intron_start, intron_end	Zero-based, half-open ASCOT exon-boundary interval.
acceptor_chrom_start/end, donor_chrom_start/end	Genomic coordinates of the two 400 nt input arms.
acceptor_window_start/end, donor_window_start/end	Segment offsets in serialized sequence.
acceptor_junction_in_sequence, donor_junction_in_sequence	Splice-junction indices in serialized sequence.
exon_id, gene_id, gene_symbol, gene_type	ASCOT/Ensembl event and gene metadata.
genome_build, sequence_matches_hg38, coordinates_source_provided	Reference-build and coordinate provenance.

Genome and annotation provenance

Sequences are extracted from the pinned UCSC hg38 FASTA and independently validated against both input-arm coordinates.

ASCOT identifies events through splice-junction evidence rather than a fixed GENCODE annotation release. The dataset therefore has no single GENCODE version defining the event set. Reported gene_id values are versioned Ensembl identifiers from the historical ASCOT/GTEx snapshot.

Processing and verification

The pinned MTSplice source, exact split membership, validation filter, sequence extraction, target equations, missing-value masks, and output hashes are embedded in the root manifest contract and documented at stable repository links:

• Processing repository: https://github.com/TaykhoomDalal/GLM-Tang-Processing/tree/main
• Task documentation: https://github.com/TaykhoomDalal/GLM-Tang-Processing/blob/main/alternative-splicing/README.md
• Reproducibility manifest: https://github.com/TaykhoomDalal/GLM-Tang-Processing/blob/main/manifest.json

Sources and citation

Tang, Z. et al. Genome Biology 26, 203 (2025). https://doi.org/10.1186/s13059-025-03674-8

Ling, J. P. et al. ASCOT identifies key regulators of neuronal subtype-specific splicing. Nature Communications 11, 137 (2020). https://doi.org/10.1038/s41467-019-14020-5

Cheng, J., Çelik, M. H., Kundaje, A. & Gagneur, J. MTSplice predicts effects of genetic variants on tissue-specific splicing. Genome Biology 22 (2021). https://doi.org/10.1186/s13059-021-02273-7

License

The MTSplice repository is MIT licensed. This dataset does not assign a new license to the historical ASCOT/GTEx measurements. Users must follow the applicable ASCOT, GTEx, and source-publication terms.