IRG Experiment Run Records (notebooks 01-15)
Archive of the experiment records of the Interpretable Regulatory Genomics project: classification of ENCODE SCREEN v4 candidate cis-regulatory elements (cCREs; PLS, pELS, dELS) in three biosamples (naive CD8+ T cell, K562, GM12878) with k-mer logistic regression / random forest / XGBoost and two small CNNs, with a focus on shortcut learning (GC, CpG and repeat composition) and interpretability. Every experiment ran as a Kaggle notebook; this dataset keeps the evidence for each run. Companion repositories: data PxALabs/encode-regulatory-elements, models irg-cnn-baselines, irg-xgboost-baselines, irg-xgboost-wide, irg-xgboost-longk, annotation irg-repeat-coverage.
This is an archive of heterogeneous files, not a single table, so the Hugging Face dataset viewer is disabled. Browse the file tree instead.
Repository layout
runs/<date>_irg-NN-<name>/
run.yaml kernel id and version, inputs, seeds, accelerator, runtime, notes
<kernel>.log Kaggle execution log (stdout)
results/ result tables (CSV), protocol.json, and for some runs JSON / NPZ / Parquet
results_docs/<name>.md written analysis of each run: findings, caveats, "what this does not show"
MANIFEST.sha256 SHA-256 of every file in the repository
README.md this card
Contents: 15 run folders (notebook numbers 01-15), 13 results documents, about 48 CSV, 17 JSON, 15 logs, 15 run.yaml, 6 Parquet and 1 NPZ file, 16 MB in total.
Runs
Dates are the Kaggle run dates (all 2026). "kernel v" is the Kaggle kernel version that produced the recorded result (earlier versions were smoke tests unless a run's run.yaml says otherwise). Runtime is wall-clock seconds on Kaggle.
| run | notebook | kernel v | runtime (s) | hardware | what was done | main finding |
|---|---|---|---|---|---|---|
20261118_irg-01-data-curation |
01 | 4 | 6,466 | CPU | Built the 9 curated tables (3 biosamples x 3 negative strategies), splits, validation report | Legacy "K562" tables were T-cell elements (ENCFF726XBE) |
20261119_irg-02-package-dataset |
02 | 1 | n/a | CPU | Verified input hashes, deduplicated, packaged v2.1 | All input hashes verified; 1-12 duplicates removed per table |
20261119_irg-03-classical-baselines |
03 | 2 | 1,501 | GPU | LR / RF / XGBoost, k-mer sweep, cross-negative and cross-cell-type tests | Random negatives inflate AUROC (0.93-0.96 vs 0.82-0.87 GC-matched); the GC-trap gate fails |
20261119_irg-04-transfer-by-class |
04 | 1 | 577 | GPU | Cross-biosample transfer by element class | Class mix explains 30-50% of the shared-vs-specific gap; about 70% remains within class |
20261119_irg-05-cpg-ablation |
05 | 1 | 705 | GPU | CpG ablation; GC+CpG-matched evaluation | Signal is distributed: removing CG k-mers costs 0.007-0.021; GC+CpG matching lowers AUROC 0.09-0.17 |
20261119_irg-06-cnn-baselines |
06 | 1 | 3,802 | GPU | ShallowCNN / AttnCNN vs k-mer XGBoost | Superseded by 07 (all runs hit the epoch cap) |
20261119_irg-07-cnn-tuned |
07 | 2 | 11,663 | GPU | Tuned CNNs: learning-rate sweep, longer training; 30 weights saved | ShallowCNN +0.007-0.012 over 06; AttnCNN mixed; advantage was measured against untuned XGBoost |
20261119_irg-08-repeat-control |
08 | 2 | 481 | GPU | Repeat-content (Alu / L1 / simple) control on top of GC+CpG matching | Repeat matching lowers AUROC only 0.010-0.025; model ordering unchanged |
20261119_irg-09-xgb-tuned |
09 | 2 | 6,248 | GPU | 24-configuration tuned XGBoost (6-mers) vs CNNs | Tuning +0.005-0.008; CNN advantage mostly gone; GM12878 stays ahead |
20261119_irg-10-motif-fdr |
10 | 2 | 4,007 | GPU + CPU | CNN first-layer filters matched to JASPAR 2024 with a shuffle null and Benjamini-Hochberg FDR (own matcher, not Tomtom) | 0.4% of 3,699 filters match at FDR 5% |
20261120_irg-11-xgb-wide |
11 | 2 | 5,725 | GPU | Wider XGBoost search, tree cap lifted to 8,000 (3 + 8 configurations per biosample) | +0.001-0.003 full AUROC over 09; XGBoost level with or ahead of ShallowCNN except GM12878 matched sets |
20261119_irg-12-multiclass-shap |
12 | 2 | 115 | GPU | Four-class (background / PLS / pELS / dELS) XGBoost with TreeSHAP | Composition explains most of the promoter advantage; drivers form a CpG gradient PLS > pELS > dELS |
20261120_irg-13-wave1-uncertainty |
13 | 2 | 629 | GPU | Re-scored all 75 saved models: paired gaps with window / chromosome-cluster / leave-one-chromosome-out intervals; negatives that are cCREs in another biosample | GM12878 CNN edge is real beyond test-sampling noise (ensembles); 5-10% of negatives are cCREs elsewhere and score higher beyond composition |
20261120_irg-14-shuffle-ladder |
14 | 2 | 1,284 | GPU | Saved models scored on k-let (k = 6..1) and block (100 / 250 bp) shuffled test windows | CNN edge survives block shuffles, not k-let shuffles: local (~7-100 bp) context; precommitted rule judged flawed and reported literally |
20261120_irg-15-longk-xgb |
15 | 2 | 8,186 | GPU | XGBoost trained on 8-mer and spaced k-mer features; same shuffle ladder | Hypothesis H0: 8-mer / spaced features do not reproduce the CNN edge and are worse than 6-mers |
Data used by the runs
- ENCODE SCREEN v4 cCREs (GRCh38), 1,000 bp windows centred on elements; biosamples
tcell,k562,gm12878; three negative strategies (gc_matchedprimary,random_unmatched,flanking). Tables:PxALabs/encode-regulatory-elementsv2.1. - Chromosome split: train chr1-14, X, Y; validation chr15-18; test chr19-22. Test chromosomes are never used for model selection.
- Test sets: full; non-homologous (8-mer Jaccard flag < 0.3); GC+CpG-matched (8 x 8 quantile bins, equal positives and negatives per cell); GC+CpG+repeat-matched (6 x 6 x 6). Test windows: T cell 12,422; K562 15,863; GM12878 16,600.
- Metrics are AUROC (and AUPRC where stated). Models are k-mer (k = 6 unless stated) XGBoost / LR / RF, ShallowCNN (7,937 parameters) and AttnCNN (189,185 parameters), 5 seeds each unless a run says otherwise. No pretrained or language models were used.
Files of note
runs/*/run.yaml: provenance of each run (Kaggle kernelpurvanshj009/irg-NN-*, version, inputs, seeds, runtime). The notebook source is in the GitHub repository underkaggle/kernels/, pinned bykernel_version.runs/*/results/protocol.json: the exact protocol (configurations, seeds, selection rules, library versions) where a run recorded one.runs/20261119_irg-10-motif-fdr/results/filter_pfm_counts.npz: filter position-frequency counts used for the motif matching;as_produced/holds the files as the kernel wrote them (see Notes).runs/20261120_irg-13-wave1-uncertainty/results/preds_<biosample>.parquet: per-window test predictions, one column per model and seed.runs/20261120_irg-14-shuffle-ladder/results/ladder_preds_<biosample>.parquet: per-level ensemble predictions on shuffled test windows.results_docs/*.md: read these before quoting any number; each states what the run does and does not show.
Notes and known issues
- The executed notebook files are not included: the Kaggle CLI exposes only the execution log and the files written to
/kaggle/working. - Early logs from the project's
experiments/logs/are unreliable and are not included. - Numbers here come from the v2.1 tables. The project's earlier tables named
k562_*were T-cell elements. - Run 10:
motif_stability.csvandmotif_jaccard_between_biosamples.csvwere recomputed frommotif_pairs_q05.csvafter a denominator bug in the kernel output; the as-produced files are inresults/as_produced/. - Run 07 weights are in
PxALabs/irg-cnn-baselines; run 09, 11 and 15 models are inirg-xgboost-baselines,irg-xgboost-wideandirg-xgboost-longkrespectively. Runs 03-06, 08, 10 and 12 did not save models. - Run 15 per-window predictions (
longk_preds_<biosample>.parquet) are only in the Kaggle kernel output, not in this dataset. - Intervals reported in the results documents cover test-sampling noise only; training-seed variance and the small number of test chromosomes (four) are stated limitations.
Integrity and reproducibility
MANIFEST.sha256 lists a SHA-256 for every file. To reproduce a run, use the notebook at the pinned kernel_version and the attached inputs named in run.yaml.
Licence and citation
Licence is other until the project owner sets one. The underlying cCRE annotation is from ENCODE (SCREEN v4); consult ENCODE's data-use policy. A manuscript is in preparation; its citation will be added here.
- Downloads last month
- 24