--- pretty_name: "GLM-Tang Task 5: INSERT-seq RNA Output" license: other tags: - biology - genomics - rna - transcription - reporter-assay - regression - benchmark - parquet configs: - config_name: archive data_files: - split: train path: archive/train.parquet - split: validation path: archive/validation.parquet - split: test path: archive/test.parquet - config_name: corrected data_files: - split: train path: corrected/train.parquet - split: validation path: corrected/validation.parquet - split: test path: corrected/test.parquet --- # GLM-Tang Task 5: INSERT-seq RNA Output This dataset contains the reporter task used in Task 5 of Tang et al. Given a 173 nt transcribed insert, predict normalized total RNA output in mouse embryonic stem cells. ## Loading ```python from datasets import load_dataset dataset = load_dataset("Taykhoom/insert-seq-glm-tang", "corrected") train = dataset["train"] validation = dataset["validation"] test = dataset["test"] ``` ## Archive dataset size | Split | Rows | |---|---:| | `train` | 9,131 | | `validation` | 1,149 | | `test` | 1,137 | | **Total** | **11,417** | The Tang Results text reports 10,774 sequences, but the released HDF5 and Methods split counts contain 11,417. This release follows the distributed HDF5. ## Sequence derivation The public GEO FASTA contains 16,461 unique 175 nt records. Removing one base from each end gives the exact 173 nt HDF5 sequence: ```python sequence = geo_sequence[1:-1] ``` Every archived sequence maps uniquely to one GEO source row. Most inserts are source-defined mm10 regions, while randomized controls are synthetic. Nullable source coordinates and explicit provenance distinguish the two cases. ## Target contract Direct row-wise matching to the GEO normalized table identifies the three HDF5 target columns: 1. `totalRNA_average` 2. `runonRNA_average` 3. `sortseq_average` The released model does **not** train on raw total RNA. It uses: ```python label = np.log(totalRNA_average + np.float32(1)) ``` This float32 expression is the bit-exact model-facing `label`. The raw and auxiliary assay values are retained as: - `raw_total_rna` - `aux_target_run_on_rna` - `aux_target_sort_seq` These are targets/measurements, not sequence features. ## Columns | Column | Description | |---|---| | `split` | Selected config's `train`, `val`, or `test` partition. | | `archive_split` | Original partition; present in `corrected` only. | | `sequence` | Exact 173 nt archived model input in transcribed orientation. | | `label` | Bit-exact float32 `np.log(totalRNA_average + np.float32(1))`. | | `example_id` | GEO FASTA sequence identifier. | | `source_index` | Zero-based row index in the 16,461-row GEO FASTA/table. | | `raw_total_rna` | Raw `totalRNA_average` assay value. | | `aux_target_run_on_rna` | `runonRNA_average`. | | `aux_target_sort_seq` | `sortseq_average`. | | `source_category`, `source_class`, `source_window`, `source_barcode` | Supplementary Table 2 design provenance. | | `locus_group_id`, `archive_locus_split_leakage` | Source `name` (unique-ID fallback for unnamed rows) and whether that group spans archive partitions. | | `chrom`, `start`, `end`, `strand`, `genome_build` | Nullable source-provided zero-based, half-open GRCm38/mm10 coordinates. | | `coordinates_available` | Whether all four source coordinate fields are present. | | `species` | Always `Mus musculus`. | | `cell_type` | Always `mouse embryonic stem cell`. | | `is_synthetic` | True only for `randomized_control` rows. | The released code trains with mean-squared error and reports test Pearson correlation and MSE on `label`. The previous `log1p` implementation differed at the bit level for 4,302 rows and is not retained as a duplicate target. ## Reproduction boundary The public source has 16,461 rows, while Tang's archive retains 11,417. The archive split is deterministic: within each full Supplementary Table 2 `class` in source order, the first `ceil(n/10)` rows are test, the next `ceil(n/10)` validation, and the remainder train; all `160-333` rows are then forced to train. Thus all 1,721 retained `160-333` rows are train-only. The `archive` config preserves those exact rows and splits. The `corrected` config assigns complete `locus_group_id` groups to a deterministic class/window-stratified 70/10/20 split. It contains 7,989 train, 1,144 validation, and 2,284 test rows; every class/window stratum with at least three groups appears in all three splits, and no locus group crosses partitions. All archived assay values equal an explicit float32 conversion of the GEO table values. Supplementary Table 2 provides coordinates for 10,346 rows. The remaining 1,071 rows include 1,059 randomized controls and 12 non-synthetic mRNA-end rows with no published locus. Six source intervals span 172 bases even though the corresponding assay sequence is 173 nt; these values are preserved as source provenance rather than silently adjusted. No hg38 mapping is claimed. The archived split contains 424 `locus_group_id` values in multiple partitions, affecting 848 rows. Group-aware evaluation should use the `corrected` config; `archive_locus_split_leakage` documents the original contamination. ## Processing and verification The HDF5 byte check, GEO sequence matching, exact float32 target conversion, target transformation, split-order verification, and the root manifest contract are 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/insert-seq/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 Vlaming, H., Mimoso, C. A., Field, A. R., Martin, B. J. & Adelman, K. *Screening thousands of transcribed coding and non-coding regions reveals sequence determinants of RNA polymerase II elongation potential*. Nature Structural & Molecular Biology 29, 613–620 (2022). https://doi.org/10.1038/s41594-022-00785-9 Source accession: GEO GSE178230. ## License The Tang archive is CC BY 4.0. GEO does not attach a standalone Creative Commons license to GSE178230; users must retain the accession/publication attribution and follow the source terms.