Work

Long Read RNA Sequencing in Primary Lung Cell Types Reveals Principles of Nonsense-Mediated Decay

Year: 2026

Type: dataset

Abstract: Source data for: long-read RNA sequencing in primary lung cell types reveals principles of nonsense-mediated decay This record contains the starting data for every analysis in the paper: the processed quantification, annotation and model outputs that the analysis code reads but cannot regenerate. Nothing here is an intermediate the code also writes — those are rebuilt from these files, by design. Design. Paired short-read and PacBio Iso-Seq long-read RNA sequencing of four primary human lung cell types — alveolar type 2 (AT2), large airway epithelial (LAE), fibroblast (FB) and microvascular endothelial (MV) — under SMG1 inhibition (SMG1i) versus vehicle (DMSO). 26 samples: 13 donor-matched DMSO/SMG1i pairs. Cell-type labels are the published ones throughout (AT2, LAE, FB, MV). Some analysis code uses the internal codes AT and DD for AT2 and LAE respectively and maps them at load time. How the record is packaged Seventeen files, 17.6 GB. Sizes here are decimal (MB = 10⁶ bytes), matching what Zenodo displays. The runtime environment is 16.9 GB of that — nmd_1.3.sif (8.1 GB) and the Docker archive it was built from (8.8 GB). It carries the R and Python versions the analyses ran under — see Runtime environment below. The data alone, without either, is 695 MB, and every analysis input is in that 695 MB. Download the runtime only if you want to run the code in the environment it was validated in. Zenodo lists files flat, and the five directories ship as archives. The analysis code resolves its DEPOSIT root to <deposit>/source_data, so put the twelve data objects one level down and leave the two manifests above them: mkdir -p nmd_deposit_2026/source_data mv *.csv *.gtf.gz *.zip nmd_deposit_2026/source_data/ mv MANIFEST.sha256 UPLOAD_MANIFEST.sha256 nmd_deposit_2026/ cd nmd_deposit_2026/source_data && for z in *.zip; do unzip -q "$z" && rm "$z"; done Extracting flat, without source_data/, is the one layout that fails silently. Every path resolves to nothing and the wiring check reports DOES NOT RESOLVE for DEPOSIT, SQANTI, ISOCALL and ANNOT — measured, not hypothetical. MANIFEST.sha256 stays ABOVE source_data/ because its paths are relative to it. Split by directory rather than bundled, so partial download works: the model exports are 7.1 MB on their own, and the Tan et al. tables are needed only for the Section 2 comparative reanalysis. UPLOAD_MANIFEST.sha256 checksums the data files and MANIFEST.sha256 covers the extracted contents; each archive also carries its own internal MANIFEST.sha256, which the file counts here exclude. 4.6 GB extracted (data only; the image is not extracted) — the SQANTI3 text compresses about 12×, while the Tan et al. .xlsx files are already compressed and gain almost nothing. Contents Quantification — the primary inputs file size shape what it is nmd_isocall_counts_4ct.csv 48.9 MB 614,993 × 26 Long-read isoform counts from the Isocall pipeline, one column per sample. The unfiltered isoform universe; the analysis applies its own expression filter (see below). nmd_lungcells_counts_4ct.csv 48.9 MB 614,992 × 26 Long-read isoform counts over the SQANTI3-filtered isoform set. salmon_gene_counts_4ct.csv 15.5 MB 78,899 × 26 Short-read gene-level counts (Salmon, length-scaled TPM→counts, via nf-core/rnaseq). 46,571 of the 78,899 genes carry a non-zero count in at least one sample. pheno_4ct.csv 2.3 kB 26 × 5 Sample metadata: sample_name, donor, treatment, cell_type, bam. The join key for every count matrix. Scope is native, not filtered downstream. These matrices contain exactly the 26 samples of the published analysis. Two exclusions applied upstream are already reflected here: the DO samples of donor 029T (a PCA outlier; both treatment partners removed to keep that cell type balanced — 029T's AT2, FB and MV samples are retained) and the air–liquid interface and small-airway conditions, which are outside the paper's scope. Code that consumes these files asserts the scope rather than re-filtering it, so an unexpected sample fails loudly. Transcript annotation file size what it is nmd_isocall_4ct.gtf.gz 42.8 MB Isoform models for the long-read universe; source of the transcript→gene map and exonic structure. sqanti.zip 348.6 MB (4.2 GB extracted) SQANTI3 output for the long-read transcriptome — structural classification, corrected models, and sequence. Treated as a starting point: the analysis begins from these calls and does not re-run SQANTI3. annotation.zip 7.3 MB Ensembl v115 gene- and transcript-level identifier maps (gmap_ENSGv115_*.rds, gmap_txlevel_ENSGv115_*.rds) used for symbol and biotype resolution; ensembl115_human_paralogs.tsv.gz with its README, the within-species paralog table behind the train/test leakage screen — homology is not derivable from a GTF, so it is deposited rather than recomputed; srsf8_ENST00000638426_reference.json, one reference transcript captured from the Ensembl REST API so the RBP section runs without network access; and msigdb_hallmark_H.tsv, the MSigDB HALLMARK collection (50 gene sets, 7,322 gene-set/symbol pairs, db_version 2026.1.Hs) captured so the productive-response section resolves hallmark membership offline instead of downloading MSigDB at run time. Key files inside sqanti.zip: nmd_lungcells_classification.txt (structural category per isoform), nmd_lungcells_filtered.gtf (filtered models), nmd_lungcells_corrected.cds.gff3 (CDS), nmd_lungcells_corrected.fasta (sequence). Isoform-pair analysis — the expressed universe and the mashr inputs isopair.zip, 21.2 MB, 5 files extracted. Two things the isoform-pair analysis cannot start without. expressed_isoforms_4ct.txt is the 95,623 isoform identifiers that passed expression filtering — the universe every downstream count in the paper is computed over. It is deposited rather than recomputed, and the reason is worth stating. The filter keeps an isoform whose share of its gene reaches 5% in either condition. A share is a ratio, so a gene with twenty comparable isoforms puts its members on 0.05 rather than near it: 2,567 isoforms lie within 1e-9 of that threshold and 983 compute to exactly 0.05. Which side of >= they fall on depends on the order a floating-point sum is accumulated in, which differs between platforms — measured with identical code, data and edgeR 4.8.2, macOS keeps 95,623 isoforms and Linux keeps 95,639, carrying through to 1,578 versus 1,580 gene-matched pairs. Neither is wrong, and no threshold rule reproduces the published split, because that split is not a rule. Reading this file reproduces the paper's numbers exactly on any platform. The code still recomputes the universe when the file is absent, now with a tolerance that makes the recomputation platform-independent, and prints which path it took; that path gives 95,817 isoforms, a 0.2% difference that changes no conclusion. nmd_mashr_die_{at2,lae,fb,mv}_2026.3.10.csv are the per-cell-type mashr differential-isoform tables that feed NMD classification. They were previously reachable only from the authors' own working tree, so the analysis could not be started from this record alone. Deep-learning model — weights, predictions and interpretability model.zip, 7.1 MB, 32 files extracted. The model predicts NMD susceptibility from sequence and ORF structure. Training requires a GPU cluster, so the trained weights and all derived outputs are deposited rather than expecting a reader to retrain. Filenames are the producing run's own and are not shortened: they carry the window configuration, the member seed and the evaluation split, because those are what distinguish one model's outputs from another's. file(s) what it is best_model_atg1000_stop1000_seed42.pt Trained weights (1000 nt windows around the start and stop codons, member seed 42). metrics_atg1000_stop1000_seed42_test_clean.json Held-out performance: AUC 0.9257, AUPRC 0.8175 over the 10,522-isoform test_clean split (2,405 NMD susceptible). evaluation_class: final_test — evaluated once. predictions_atg1000_stop1000_seed42_all.tsv Per-isoform predictions for all 41,776 isoforms, with h5_split (train 26,720 / val 4,356 / test 10,522 / test_paralog 122 / val_paralog 56), label, logit, prob. predictions_atg1000_stop1000_seed42_test_clean.tsv The held-out split alone (10,522 rows) — a strict subset of the above. kernel_shap_branch_atg1000_stop1000_seed42_all.tsv KernelSHAP attributions per input branch (sequence, ORF structure, stop-window), all isoforms. Exact Shapley by enumeration over three players — not the sampling estimator, so shap.KernelExplainer will not reproduce it. deepshap_summary_atg1000_stop1000_seed42_{joint,structural,atg-stop}_run{1..5}.tsv DeepSHAP feature attributions: three DISTINCT decompositions, five replicates each. The mode is in the filename deliberately — they are different analyses and must not be pooled. shap_profile_{atg,stop}_joint_atg1000_stop1000_seed42.tsv Per-nucleotide signed SHAP around the start and stop codons. motif_logo_atg_atg1000_stop1000_seed42_run1.tsv, motif_logos_stop_atg1000_stop1000_seed42_run1.tsv Position-weight matrices for the sequence-logo panels. The two differ in stem (logo / logos) because the producers do; the names are not normalized here. uorf_attention_{metrics,predictions,subgroup_counts}.tsv, uorf_features_in_priority_slots.tsv Model attention over candidate upstream ORFs, and the subgroup counts behind the uORF figure panels. The model was trained on the universe these files rebuild — 41,776 isoforms. Splits are chromosome-deterministic and sequences are byte-identical to the deposited FASTA across the whole universe. The interpretability exports are deposited as starting points: the figures and statistics in the paper's model section are computed from these tables, so they can be reproduced without a GPU. Runtime environment — the container image nmd_1.3.sif (8.1 GB) and nmd_docker_1.0_2026-08-13_verified.tar (8.8 GB) — the same environment as

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.21544336