Ickes, Christian · Cigdem Hazal, Timucin · Harms, Bendix Christian · et al.

3 files · 13 GB · gzip, zipdeclared

TFCP_model.zip Contains the official fine-tuned DNABERT model for DNA-binding domain class prediction, as described in TFClassPredict. TFCP_precompiled.zip Contains precompiled genome-wide binding potential tracks in UCSC hg38 coordinates, provided as one genomic track per DNA-binding domain class. UNIBIND_TFBS_dataset_filtered.csv.gz This dataset comprises more than 5 million high-confidence TFBS sequences derived from the UniBind database, covering 265 TFs with TFClass annotations across 23 DBD classes. All binding sites are annotated by the DBD class of the corresponding bound TF and standardized to a uniform length of 21 bp, with underlying sequences extracted from the GRCh38/hg38 reference genome. Overlapping binding sites and those located on chromosomes X and Y were excluded to reduce class assignment ambiguity and minimize sex-chromosome-specific confounding effects. The dataset is partitioned into training, validation, and test sets using a chromosome-stratified 80:20 split, ensuring consistent chromosomal representation across all partitions and a strict separation between model development and final performance evaluation.