Multi-modal transcript isoform expression dataset

We curated the human transcript isoform expression dataset from the GTEx portal following the preprocessing pipeline in Garau-Luis et al. (2024). We downloaded the RNA-seq Transcript TPMs file from the bulk tissue expression in GTEx Analysis V8. The table contains transcript expression collected from 30 non-diseased tissues in nearly 1000 human individuals. We averaged the transcript expression measurements across individuals to obtain population-level tissue-specific transcript expression data. For each transcript, we used the associated Ensembl transcript id to retrieve a DNA sequence of length 197k centering on the transcript start site using the human reference genome GRCh38. We also retrieved the protein sequence for each transcript if it encodes a protein. For each sample in the dataset, we have a DNA sequence, an RNA sequence, a protein sequence (potentially), and expression levels of the transcript across 30 tissues. Furthermore, we split the dataset into training, validation and test sets based on the chromosomes. In specific, we used chromosomes 1-19 for training, 20 for validation and 21-22 for test. The training, validation, and test sets contain 175k, 4k and 6.5k samples respectively. We performed $log_2(0.0001+x)$ transform for the labels.

Details about dataset

• training set: train_i.tsv, where i=1,...,19 denotes chromosome i.
• validation set: validation.tsv, corresponds to chromosome 20.
• test set: test.tsv, corresponds to chromosomes 21 and 22.

Columns in each file:

• transcript_id
• gene_id
• labels_i (i=0,...,29): transcript isoform expression level in 30 tissues.
• chromosome
• tss: transcription start site
• tes: transcription end site
• strand
• rna_seq
• protein_seq
• dna_seq: of length 197k, centered at the TSS of the transcript

Update Note — 2026-07-10

1. Fixed a bug in transcript coordinate annotation for genes on the negative (-) strand: the TSS should correspond to the transcript end, and the TES should correspond to the transcript start. In the previous version, these coordinates were reversed, which affected dna_seq for transcripts on the negative strand.
2. Added three new columns: coding_start, coding_end, and splice_sites.