2026-10-14T10:44:11Z https://api.figshare.com/v2/oai

oai:figshare.com:article/33411454 2026-10-01T01:24:19Z category_24310 category_24175 item_type_3 month_year_10_2026

A cross-tissue single-cell atlas of full-length human transcript isoforms: count matrices, metadata, and feature annotation Madhav Mantri (24649740) Genomics Systems biology long-read RNA-seq data Tabula Sapiens Overview: Single-cell full-length transcript isoform atlas of 203,311 cells from 26 tissues of 12 human donors, generated by PacBio Kinnex (MAS-Seq) long-read sequencing of the same barcoded cDNA used for Tabula Sapiens short-read profiling. Cells carry consensus cell-type labels (144 fine types across 34 broad classes). This deposit contains the analysis-ready count matrices, per-feature annotation, the full exon structure and splice junctions of every transcript structure, the sample table, and per-cell CDKN2A isoform calls for the four donors assayed by targeted capture.Feature levels: Counts are provided at three nested levels. Because a transcript structure and an annotated transcript are not the same object, the levels are named explicitly rather than all being called "isoforms":<ol><li>`genes` — one feature per gene (49,124 features)</li><li>`pbids` — one feature per distinct collapsed transcript structure, identified by its PacBio ID, including both annotated and novel structures (854,410 features over 37,805 genes)</li><li>`ensemblids` — pbid counts summed within each transcript identifier, so structures matching the same GENCODE transcript are pooled while structures matching no annotated transcript retain their PacBio ID (489,842 features)</li></ol>All three share an identical cell set of 203,311 cells, so they can be used interchangeably or side by side.Files:<ul><li>all_samples_pacbio_recollapsed_raw_counts_genes_bc_anndata_preprocessed.h5ad (~1.8 GB) — gene-level count matrix, 203,311 cells × 49,124 genes, AnnData/h5ad.</li><li>all_samples_pacbio_recollapsed_raw_counts_pbids_bc_anndata_preprocessed<code>.h</code>5ad (~2.4 GB) — pbid-level count matrix, 203,311 cells × 854,410 transcript structures, AnnData/h5ad.</li><li>all_samples_pacbio_recollapsed_raw_counts_ensemblids_bc_anndata_preprocessed.h5ad (~3.1 GB) — ensemblid-level count matrix, 203,311 cells × 489,842 features, AnnData/h5ad.</li><li>all_samples_pbids_feature_metadata_sqanti3.csv (~263 MB) — SQANTI3 annotation for every pbid feature, 854,410 rows × 59 columns. Includes structural category, associated gene and transcript, splice-junction and transcript-boundary evidence, CAGE and poly(A) support, and coding, CDS and predicted-NMD calls. The last two columns flag structures whose annotation is unreliable (see Notes for users).</li><li>all_samples_pbids_feature_metadata_pigeon.csv (~304 MB) — pigeon annotation for every pbid feature, 854,410 rows × 53 columns. Structural classification against GENCODE v41. The last two columns are the same reliability flags as in the SQANTI3 table.</li><li>sample_metadata.csv (8 KB) — sample table, 62 rows, mapping tube_id to donor, tissue, tube_label, long-read library ID and matched short-read sample ID.</li><li>xgen_cdkn2a_per_cell_pigeon_raw.csv (19 MB) — per-cell CDKN2A isoform counts and senescence status, 155,117 rows × 27 columns. One row per atlas cell of the four donors carrying a targeted xGen CDKN2A capture (TSP21, TSP25, TSP27, TSP33); the remaining eight donors were not assayed and are absent. Columns are the 14 annotated CDKN2A transcripts plus a single `novel` bin holding all structures that matched no annotated transcript, followed by `cdkn2a_total`, the per-class totals `class_p16` / `class_p14ARF` / `class_other_CDKN2A`, the call `cdkn2a_pos` / `cdkn2a_status`, and `xgen_matched`. Indexed by `cell_id`, matching `adata.obs_names`.</li><li>all_samples_pbids_isoform_structures.gtf.gz (~31 MB) — exon structure of every pbid feature in GTF format: one `transcript` line and its `exon` lines per pbid (854,410 transcripts, 4,531,978 exons), with `gene_id` and `transcript_id` attributes. GRCh38, 1-based closed coordinates. The transcript line gives the 5′ and 3′ ends (on the minus strand the 5′ end is the `end` column). Coordinate-sorted and bgzip-compressed; index with `tabix -p gff` for region queries.</li><li>all_samples_pbids_isoform_structures.bed.gz (~26 MB) — the same structures in BED12 format, one line per pbid with exons as blocks, for genome browsers. 0-based half-open coordinates; no CDS is given, so thickStart = thickEnd. Coordinate-sorted and bgzip-compressed; index with `tabix -p bed`.</li><li>all_samples_pbids_junctions.tsv.gz (~19 MB) — every splice junction of every pbid feature, 3,677,568 rows (one per junction per transcript; the 20,292 mono-exon pbids have none). Columns: `isoform`, `chrom`, `strand`, `junction_number` (numbered 5′ to 3′ along the transcript), `intron_start` and `intron_end` (1-based first and last intron base), `junction_category`, `donor_site_category` and `acceptor_site_category` (known or novel relative to GENCODE v41), `splice_site` (dinucleotide motif read from the GRCh38 reference) and `canonical` (GT-AG, GC-AG or AT-AC).</li><li>all_samples_pbids_within_tube_merges.csv (~8 MB) — the 156,130 pbids whose identifier covers more than one per-sample structure in at least one sample (see Notes for users). Columns: `isoform`, `n_tubes`, `max_structures_in_a_tube`, `tubes`, `merged_in_tissue_source`.</li></ul>In each h5ad, `.X` holds raw integer counts. `.obs` carries donor, tissue, tube_id, LR_library_id, SR_sample_id, popv_prediction, cell_ontology_class, broad_cell_class and per-cell QC metrics. `.var` carries gene_name, transcript_id and pigeon annotation.Joining the feature metadata: Both CSVs have exactly one row per feature of the pbid object; join on the `isoform` column against `adata.var_names`. Coverage is one-to-one and complete. The `tissue_source` column records which sample contributed a given structure — it is a per-feature label, not a detection list, and a feature may be present in many tissues.CDKN2A calls. `xgen_cdkn2a_per_cell_pigeon_raw.csv` joins on `cell_id` against `adata.obs_names`, so `adata.obs.join(df.set_index("cell_id"))` attaches it directly. `cdkn2a_status` is one of `p16+`, `p14ARF+`, `Both`, `other_CDKN2A+`, `CDKN2A-` or `not_in_pulldown`. Per-transcript counts are given as columns rather than PacBio identifiers because the four captures were collapsed independently and their PacBio namespaces are disjoint, so an identifier means a different structure in each donor; the annotated transcript set is shared and therefore comparable.Transcript structures. The GTF, BED12 and junction files describe the structure that defines each pbid identifier: the model produced by the global re-collapse across all samples. They join to the count matrices and the feature metadata on `transcript_id` (GTF), `name` (BED12) or `isoform` (junction table), all equal to `adata.var_names` of the pbid object. These files hold coordinates only, with no read sequence.Processing: Reads were processed with the PacBio Iso-Seq workflow, aligned to GRCh38.p13, and collapsed per sample. Transcript identifiers were then harmonised across all samples by a global re-collapse, so one structure has one identifier atlas-wide. Models were classified against GENCODE v41 with pigeon (pbpigeon v26.2.0); coding, CDS and predicted-NMD annotation is from SQANTI3 v6.0.1. Cell types were predicted with popV v0.6.0 and mapped to Cell Ontology terms. Cells were retained at more than 1,000 UMIs on the gene matrix; genes at detection in at least 5 cells; and individual isoform-level features at detection in at least 5 cells with at least 10 total counts across the atlas.Notes for users:<ul><li>`ORF_genomic_span` in the SQANTI3 table is the genomic span of the CDS, not the length of the ORF, so for multi-exon transcripts it includes intron content. Use `CDS_length` for coding-sequence length.</li><li>The pigeon table's coding, ORF, CDS and predicted-NMD columns are empty by design; SQANTI3 is the source for those.</li><li>`sample_metadata.csv` lists all 62 sequenced samples across 13 donors. The atlas comprises the 60 samples from 12 donors; samples T26 and T27 (donor TSP31) were excluded and do not appear in the matrices.</li><li>Use `xgen_matched` as the denominator for any CDKN2A rate. Only 8,954 of the 155,117 rows were seen by the capture; the other 146,163 carry zeros because they were never assayed, not because CDKN2A was absent, and are labelled `not_in_pulldown`.</li><li>The global re-collapse can assign one identifier to several per-sample structures that share an intron chain but differ in their ends. When that happens within the sample a feature-metadata row was taken from (its `tissue_source`), the row's SQANTI3 or pigeon annotation may mix those structures. Such rows have `structure_merged_in_tissue_source` = True (30,281 of 854,410 pbids); for them, use the structure in the GTF, BED12 and junction files rather than the exon count, length or coordinates in the metadata row. `n_tubes_structure_merged` counts the samples in which the identifier covers more than one structure.</li></ul>Associated code: https://github.com/madhavmantri/tabula_longread 2026-09-01T00:00:00Z Dataset Dataset 10.6084/m9.figshare.33411454.v2 https://figshare.com/articles/dataset/A_cross-tissue_single-cell_atlas_of_full-length_human_transcript_isoforms_count_matrices_metadata_and_feature_annotation/33411454 CC BY 4.0