IRG XGBoost on longer and spaced k-mers (notebook 15)
Twenty-one gradient-boosted classifiers that predict whether a 1,000 bp window is an ENCODE candidate cis-regulatory element (cCRE: PLS, pELS or dELS) of one biosample, against GC-matched genomic background. They test whether k-mer features longer or gapped than the 6-mers of PxALabs/irg-xgboost-wide can reproduce the small CNN advantage seen on GM12878. They do not: they are less accurate than the 6-mer models. Published as a documented negative result, not as recommended models. Part of github.com/PxA-Labs/interpretable-regulatory-genomics. No pretrained or language models are involved.
Files
models/<biosample>_xgb_<features>_s<seed>.ubj- XGBoost native binary models (no pickle); biosampletcell,k562,gm12878;<features>isk8: 8-mer frequencies, 65,536 sparse features (seeds 0-2),spaced: three weight-6 gapped k-mers (offsets 0,1,3,4,6,7 / 0,1,4,6,7,8 / 0,2,3,5,6,8), 3 x 4,096 sparse features (seeds 0-2),k6_ctrl: 6-mer frequencies through the same sparse pipeline (seed 0 only; a pipeline control).
longk_features.py- the feature functions (sparse_feats,MASKS). Features are frequencies of the k-mers present; absent k-mers are zeros in CSR (treated as missing by XGBoost).results/- search, final-model, ladder, gap and diagnostics tables,protocol.json;MANIFEST.sha256- SHA-256 of every model file.
Usage
import xgboost as xgb
from longk_features import sparse_feats, MASKS
clf = xgb.XGBClassifier(); clf.load_model("models/gm12878_xgb_k8_s0.ubj")
p = clf.predict_proba(sparse_feats(["ACGT" * 250], MASKS["k8"]))[:, 1]
Use all seeds and report the spread; do not pick a seed on test data.
Training
Same data, split (train chr1-14,X,Y / validation chr15-18 / test chr19-22) and test sets as notebooks 07-14. Per biosample and feature set two configurations were tried: the notebook-11 selected configuration (C0) and the same with colsample_bytree halved (C1); the better on validation AUROC was used for the final seeds (test chromosomes never used for selection). Up to 6,000 trees, early stopping 100, max_bin 64. Source: Kaggle kernel purvanshj009/irg-15-longk-xgb v2, XGBoost 3.4.1 on GPU.
Evaluation (test chromosomes, ensemble AUROC of the seed-averaged probability, 3 seeds)
| biosample | 6-mer (nb11) full / matched | k8 full / matched | spaced full / matched |
|---|---|---|---|
| tcell | 0.893 / 0.749 | 0.879 / 0.740 | 0.881 / 0.714 |
| k562 | 0.873 / 0.776 | 0.858 / 0.762 | 0.856 / 0.740 |
| gm12878 | 0.848 / 0.760 | 0.839 / 0.753 | 0.838 / 0.739 |
"Matched" = test subset matched on GC and CpG observed/expected, equal positives and negatives per bin (a sub-population, not comparable with the full-task AUROC). Against the notebook-07 CNN ensembles on GM12878 (matched) the validation-selected k8 model is behind by 0.018 (ShallowCNN, 95% CI +0.012 to +0.024) and 0.025 (AttnCNN, +0.019 to +0.031): the CNN edge is not reproduced by these features. Full analysis and caveats: docs/results/longk_xgb.md in the GitHub repository.
Intended use and limitations
- Research artefacts documenting a negative result. Not for clinical use or variant-effect prediction; use
PxALabs/irg-xgboost-widefor a k-mer baseline. - Each model is specific to the biosample it was trained on. Negatives are GC-matched genomic windows outside the biosample's cCREs, not validated non-regulatory elements.
- Only two feature sets and a two-configuration search were tried; this does not show that no k-mer model can match the CNN.
- The
k6_ctrlmodels, which use the same sparse pipeline as the new features, score 0.0015-0.0047 below the notebook-11 dense 6-mer models on the full test set (seed 0), so part of the deficit of k8/spaced is pipeline, not features. - Three seeds; four test chromosomes; intervals cover test-sampling noise only.
- Data note: the project's earlier tables named
k562_*were T-cell elements; these models use the corrected v2.1 tables.
Provenance and integrity
The 21 files were uploaded from inside Kaggle (kernel irg-15b-push-longk-to-hf); SHA-256 is in MANIFEST.sha256. Not yet done for this repo: re-scoring the models on the test rows outside the training run (as was done for irg-xgboost-wide); the AUROCs above come from the training run itself.