Dataset: Interpretable multimodal learning from sequence and genomic context for lncRNA classification

This dataset contains all data files and experimental results associated with the manuscript "Interpretable multimodal learning from sequence and
genomic context for lncRNA classification".

Related Code Repository: https://github.com/cbib/beta_vae_lnclassifier

Code Archive DOI: 10.5281/zenodo.18833347

Manuscript: [Citation when available]

File Organization

The dataset is organized into three ZIP archives:

1. data.zip - Input data files and preprocessing outputs
2. gencode_v47_experiments.zip - All experimental results on GENCODE v47
3. gencode_v49_experiments.zip - All experimental results on GENCODE v49

1. data.zip

Contains all input sequences, features, and lncRNA-BERT baseline results.

Contents:

cdhit_clusters/

CD-HIT clustered transcript sequences for training:

• v47_lncRNA_clustered.fa - CD-HIT clustered lncRNA sequences (v47)
• v47_pc_clustered.fa - CD-HIT clustered protein-coding sequences (v47)
• v49_lncRNA_clustered.fa - CD-HIT clustered lncRNA sequences (v49)
• v49_pc_clustered.fa - CD-HIT clustered protein-coding sequences (v49)

dataset_biotypes/

Biotype annotations for datasets:

• v47_dataset_biotypes_cdhit.csv - Transcript biotype labels (v47)
• v49_dataset_biotypes_cdhit.csv - Transcript biotype labels (v49)

Format: CSV with columns including transcript_id, biotype, gene_id

lncRNABERT_results/

Zero-shot baseline results from lncRNA-BERT:

• v47_lncRNABERT_embeddings.h5 - Learned embeddings (v47)
• v47_lncRNABERT_results.csv - Predictions and metrics (v47)
• v49_lncRNABERT_embeddings.h5 - Learned embeddings (v49)
• v49_lncRNABERT_results.csv - Predictions and metrics (v49)

processed_features/

Cleaned and normalized feature vectors with associated metadata:

• v47_nonb_feature_names.txt - Non-B DNA feature names (v47)
• v47_nonb_features_clean.csv - Processed non-B DNA features (v47)
• v47_nonb_scaler.pkl - Scikit-learn scaler for non-B features (v47)
• v47_te_feature_names.txt - TE feature names (v47)
• v47_te_features_clean.csv - Processed TE features (v47)
• v47_te_scaler.pkl - Scikit-learn scaler for TE features (v47)
• v49_nonb_feature_names.txt - Non-B DNA feature names (v49)
• v49_nonb_features_clean.csv - Processed non-B DNA features (v49)
• v49_nonb_scaler.pkl - Scikit-learn scaler for non-B features (v49)
• v49_te_feature_names.txt - TE feature names (v49)
• v49_te_features_clean.csv - Processed TE features (v49)
• v49_te_scaler.pkl - Scikit-learn scaler for TE features (v49)

Description: Feature scalers (.pkl) can be loaded with scikit-learn to apply the same normalization used during training.

split_gencode_47/

Train/test split for GENCODE v47:

• lnc_test.fa - lncRNA test set
• lnc_trainval.fa - lncRNA training+validation set
• pc_test.fa - Protein-coding test set
• pc_trainval.fa - Protein-coding training+validation set
• split_manifest.json - Split metadata and statistics

split_gencode_49/

Train/test split for GENCODE v49 (same structure as split_gencode_47/)

2. gencode_v47_experiments.zip

Experimental results for all models trained and evaluated on GENCODE v47.

Contents:

beta_vae_contrastive_g47/

β-VAE with contrastive learning (sequence-only baseline):

• evaluation_csvs/ - Evaluation metrics and predictions
• global_biotype_enrichment/ - Biotype enrichment analysis
• models/ - Model checkpoints
• performance_figures/ - Performance visualization plots
• spatial_analysis/ - Spatial clustering analysis
• umap_visualizations/ - UMAP embedding visualizations
• ANALYSIS_SUMMARY.md - Summary of key findings
• biotype_mapping.json - Biotype label mappings
• cv_evaluation_results.json - Cross-validation results
• cv_fold_results.csv - Per-fold cross-validation metrics
• embeddings_all_folds.npz - Concatenated embeddings from all CV folds
• embeddings_best_fold.npz - Embeddings from best performing fold
• model_architecture.txt - Model architecture description
• model_paths.csv - Paths to saved model files
• test_results.json - Final test set results

beta_vae_features_attn_g47/

β-VAE with attention-based feature fusion (TE + non-B DNA): Same structure as beta_vae_contrastive_g47/

beta_vae_features_g47/

β-VAE with concatenated features (TE + non-B DNA): Same structure as beta_vae_contrastive_g47/

cnn_g47/

CNN baseline (sequence-only): Same structure as beta_vae_contrastive_g47/

stat_results/

Statistical analysis results across all models:

ablations_v47/

Ablation study results:

• bootstrap_f1_ci.csv - Bootstrap confidence intervals for F1 scores
• delongauc_ci.csv - DeLong test for AUC comparisons
• fold_summary.csv - Summary statistics per fold

g47/

GENCODE v47 statistical analysis:

• g47_bootstrap_f1_ci.csv - Bootstrap F1 confidence intervals
• g47_fold_summary.csv - Per-fold summary statistics
• hardcase_jaccard_pairwise_v47.csv - Jaccard similarity for hard cases
• hardcase_jaccard_v47.csv - Hard case Jaccard indices
• hardcase_membership_long_v47.csv - Hard case membership matrix
• hardcase_upset_v47.png - UpSet plot for hard case overlaps

3. gencode_v49_experiments.zip

Experimental results for all models trained and evaluated on GENCODE v49.

Contents:

Same directory structure as gencode_v47_experiments.zip:

• beta_vae_contrastive_g49/
• beta_vae_features_attn_g49/
• beta_vae_features_g49/
• cnn_g49/
• stat_results/ablations_v49/ and stat_results/g49/

Reproducibility

To reproduce the results: Refer to the code repository (DOI: 10.5281/zenodo.18833347) for scripts

The split_manifest.json files document the exact train/test splits used.

Citation

If you use this dataset, please cite:

[Author list]. (2026). [Manuscript title]. 
Bioinformatics. DOI: [DOI when available]

Dataset DOI: 10.5281/zenodo.18849718
Code DOI: 10.5281/zenodo.18833347

License

 CC BY 4.0

Contact

For questions or issues regarding this dataset, please contact:

• Mikaël Georges: mikael.georges@ibgc.cnrs.fr | Macha Nikolski macha.nikolski@u-bordeaux.fr
• Or open an issue on the GitHub repository: https://github.com/cbib/beta_vae_lnclassifier

Last Updated: 03/03/26
Version: 1.0.0