DELDA: pre-trained checkpoints and processed dataset for lncRNA–disease association prediction

This record archives the trained model checkpoints and processed dataset accompanying the paper"DELDA: Dual-Encoder Co-Attention with Large-Scale Biomedical Pre-trained Models forlncRNA–Disease Association Prediction" (Sanguk Kim & Jihwan Ha, Pukyong National University). DELDA predicts long non-coding RNA (lncRNA)–disease associations directly from raw nucleotidesequences and disease literature, combining a CNN-BERT lncRNA encoder, a BioBERT diseaseencoder, and a bidirectional co-attention classifier — without relying on association-networktopology. CONTENTS Model checkpoints- best_model_fold_0.pt — CNN-BERT lncRNA encoder, pre-trained by masked language modeling on 173,524 Homo sapiens lncRNA sequences (6-mer tokenization). Required to (re)train any DELDA variant.- full_training_model.pt — Complete DELDA model trained on the full dataset; used for single-pair inference (inference.py).- base_fold1_best.pt — DELDA weights from fold 1 of the 5-fold cross-validation; the exact model used for the case study (manuscript Tables 8–9). Processed dataset (place in the code's data/ directory)- positive_associations.csv (1,519), negative_associations.csv (1,519), unlabeled_associations.csv (13,318) — lncRNA–disease pairs.- LncRNA_indices_mapped.csv (89) — lncRNA → nucleotide sequence.- Disease_indices_mapped.csv (190) — disease → MeSH term + definition. Pre-training corpus- sequences_str_list.npz — the 173,524 Homo sapiens lncRNA sequences used for MLM pre-training. Not required for downstream training or inference (best_model_fold_0.pt already encapsulates it); provided only to reproduce the pre-training stage from scratch. SOURCES AND LICENSING Processed data are derived from MNDR (associations), RNAcentral (sequences), and MeSH/NLM(disease text). Lnc2Cancer 3.0 and LncRNADisease v3.0 were used only for external validation inthe case study and are not redistributed here. Because the association data inherit anon-commercial restriction from MNDR, all files in this record are released under CC BY-NC 4.0,with one exception: sequences_str_list.npz is pure RNAcentral content and is therefore CC0 1.0.The accompanying source code is distributed separately under the MIT license. Please cite MNDR,RNAcentral, and MeSH in any downstream use.