Uncertainty-aware benchmarking reveals ambiguous transcripts in mRNA–lncRNA classification

Abstract

Zenodo data README This dataset contains data files and experimental results associated with the manuscript "Uncertainty-aware benchmarking reveals ambiguous transcripts in mRNA–lncRNA classification". Related Code Repository: https://github.com/cbib/uncertainty_aware_lnclassifier Data Archive DOI: 10.5281/zenodo.19551649 Manuscript: [Citation when available] File organization The attached data.zip contains the following sections and files: Features gencode.v47.common.cdhit.cv_full_table.tsv — Complete feature matrix: all raw scores and computed features from every classifier, per transcript gencode.v47.common.cdhit.cv_binary_class_table.tsv — Binary classification matrix: classification assigned by each of the tools, per transcript. (lncRNA = False/protein-coding = True). gencode.v47.common.cdhit.cv_probs_table.tsv — Coding-probability scores (0–1) from each classifier per transcript, used as input for ensemble/entropy analysis Entropy gencode.v47.common.cdhit.cv_entropy_groups.tsv — Low / middle / high entropy group assignment for each transcript (two-column: seq_ID, entropy_group) gencode.v47.common.cdhit.cv_uncertainty_analysis.tsv — Per-transcript entropy metrics (H_pred, I_bald) plus per-tool entropies, true class, and biotype Stats high_entropy_pc_v_lnc_cat_freq.tsv — Categorical feature frequencies (group1 vs group2) for high-entropy coding vs lncRNA transcripts high_entropy_pc_v_lnc_chi2.csv — Chi-squared + Cramér's V + odds-ratio results for categorical features, high-entropy coding vs lncRNA high_entropy_pc_v_lnc_mannwhitney.csv — Mann-Whitney U + VDA results for continuous features, high-entropy coding vs lncRNA low_entropy_pc_v_lnc_cat_freq.tsv — Categorical feature frequencies for low-entropy coding vs lncRNA transcripts low_entropy_pc_v_lnc_chi2.csv — Chi-squared + Cramér's V + odds-ratio results for categorical features, low-entropy coding vs lncRNA low_entropy_pc_v_lnc_mannwhitney.csv — Mann-Whitney U + VDA results for continuous features, low-entropy coding vs lncRNA low_vs_high_entropy_cat_freq.tsv — Categorical feature frequencies comparing all low-entropy vs high-entropy transcripts low_vs_high_entropy_chi2.csv — Chi-squared + Cramér's V + odds-ratio results for low vs high entropy (all transcripts) low_vs_high_entropy_mannwhitney.csv — Mann-Whitney U + VDA results for low vs high entropy (all transcripts) supp_coding_low_vs_high_entropy_cat_freq.tsv — Categorical feature frequencies for low vs high entropy within coding transcripts only supp_coding_low_vs_high_entropy_chi2.csv — Chi-squared results for low vs high entropy, coding-only supp_coding_low_vs_high_entropy_mannwhitney.csv — Mann-Whitney U + VDA results for low vs high entropy, coding-only supp_lncrna_low_vs_high_entropy_cat_freq.tsv — Categorical feature frequencies for low vs high entropy within lncRNA transcripts only supp_lncrna_low_vs_high_entropy_chi2.csv — Chi-squared results for low vs high entropy, lncRNA-only supp_lncrna_low_vs_high_entropy_mannwhitney.csv — Mann-Whitney U + VDA results for low vs high entropy, lncRNA-only top_features_from_clusters.csv — Top discriminative features selected after FDR correction and cluster-based deduplication, with VDA and test statistics Clustering feature_correlation_matrix.csv — Full pairwise Spearman correlation matrix across all features feature_clusters_at_distances.csv — Cluster ID assigned to each feature at every correlation-distance threshold (0.05–1.45 step 0.05) silhouette_scores.csv — Silhouette score and number of clusters at each distance threshold optimal_threshold.txt — Distance threshold that maximises silhouette score feature_correlation_dendrogram.pdf — Hierarchical clustering dendrogram of pairwise Spearman feature correlations SHAP performance_summary.csv — Precision / recall / F1 / accuracy summary across CV folds for the random-forest model all_predictions.csv — Per-transcript, per-fold random-forest classification predictions (true label, predicted label, class probabilities) shap_aggregated.csv — Mean ± SD absolute SHAP value per feature, aggregated across all transcripts and folds shap_all_transcripts.csv — Raw SHAP values for every transcript × feature × fold shap_per_fold_mean_abs.csv — Mean absolute SHAP value per feature per CV fold REP features (RNA) all_transcripts.out — Raw RepeatMasker output for all GENCODE v47 transcripts (Smith-Waterman scores, divergence, repeat family/class) all_transcripts_te_features.csv — Repetitive element summary features per transcript derived from RepeatMasker (counts, scores, coverage by TE class etc.) REP features (DNA) all_transcripts_te_features.csv — RepeatMasker summary features from unspliced genomic sequences. The REP features (RNA) section contains the separate spliced RNA results. NBD features features_nonb_features.csv — Non-B DNA structural features per transcript (A-phased repeats, G-quadruplexes, Z-DNA, etc.) ScanFold scanfold_final_partners_stats.tsv — RNA secondary-structure summary per transcript. rG4 peak_counts.csv — Predicted RNA G-quadruplex peak counts per transcript. Configuration config.yaml, feature_analysis_config.yaml, shap_config.yaml, figures_config.yaml, shap_cherry_picks.json — Analysis, feature, SHAP and figure settings. Snakevision pipeline graphs nonb-pipeline-dag.dot, nonb-pipeline-dag.svg, revision-wt-dag.dot, revision-wt-dag.svg, rulegraph-dashboard.svg, secondrna-dag.dot, secondrna-dag.svg, te-pipeline-dag.dot, te-pipeline-dag.svg — Pipeline diagrams in DOT and SVG formats. Archive layout and restoration manifest.json maps archive members to pipeline-relative paths. Sections keep files with the same basename separate, including RNA and DNA repeat features and full and clustered SHAP results. To restore the archive into a pipeline checkout, use the companion script: python unzip_zenodo.py --zip /path/to/data.zip --paper-dir /path/to/checkout --dry-run python unzip_zenodo.py --zip /path/to/data.zip --paper-dir /path/to/checkout The second command writes files and replaces existing files at mapped paths.