TransBind2: Improving Transcription Factor-DNA Binding Prediction with Multimodal Data and Bidirectional Cross Attention
- 1University of Missouri - Columbia
Abstract
Accurate genome-wide prediction of transcription factor (TF)-DNA binding remains challenging because many models focus mainly on DNA sequence and overlook chromatin context and TF structure. We previously developed TransBind, a protein-aware model that combines TF and DNA representations through cross-attention. Here, we introduce TransBind2, which improves on TransBind in several ways. It incorporates DNase-seq accessibility and genome mappability tracks as additional input, uses a biomodal protein language model (ProstT5) to capture both TF sequence and structure, and applies bidirectional cross-attention so DNA and protein features can refine each other. We also frame prediction as binary classification of individual triplets, allowing the model to generalize to new TFs and cell types. Across 690 human ChIP-seq experiments covering 161 TFs and 91 cell types, TransBind2 achieves a macro AUROC of 0.9648 and AUPR of 0.4215, outperforming TransBind and other baselines, with a [≥]12.67% relative AUPR gain. The model trained on human data also performs well in cross-species zero-shot prediction on mouse data. Saliency analysis shows that it can identify TF-binding peaks with a median error of 12-38 base pairs (bps) despite being trained on window-level labels. Ablation studies further show that TF structure, chromatin accessibility, and bidirectional attention each improve performance. Overall, these results show that combining TF structure with chromatin context leads to more accurate and generalizable TF-DNA binding predictions.
Related
- Double-Take: Improving Capacity, Binding, and Selectivity In Long-Context Within Bidirectional Memory
- Aging reshapes cardiac extravascular leukocyte transcription through bidirectional remodeling during LPS-induced endotoxemia
- GraphTFactor: A graph-based approach for proteins transcription factor prediction
- EPIC: An open community challenge for sequence-based prediction of transcription initiation in five non-model metazoans
- CaliPPer: quantifying, predicting and improving AI model performance for binding prediction
BibTeX
@misc{basnet2026transbind2,
title = {{TransBind2: Improving Transcription Factor-DNA Binding Prediction with Multimodal Data and Bidirectional Cross Attention}},
author = {S. Basnet and J. Cheng},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.07.749913},
url = {https://www.biorxiv.org/content/10.64898/2026.09.07.749913v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - TransBind2: Improving Transcription Factor-DNA Binding Prediction with Multimodal Data and Bidirectional Cross Attention
AU - S. Basnet
AU - J. Cheng
PY - 2026
DA - 2026/09/07
PB - bioRxiv
DO - 10.64898/2026.09.07.749913
UR - https://www.biorxiv.org/content/10.64898/2026.09.07.749913v1
AB - Accurate genome-wide prediction of transcription factor (TF)-DNA binding remains challenging because many models focus mainly on DNA sequence and overlook chromatin context and TF structure. We previously developed TransBind, a protein-aware model that combines TF and DNA representations through cross-attention. Here, we introduce TransBind2, which improves on TransBind in several ways. It incorporates DNase-seq accessibility and genome mappability tracks as additional input, uses a biomodal protein language model (ProstT5) to capture both TF sequence and structure, and applies bidirectional cross-attention so DNA and protein features can refine each other. We also frame prediction as binary classification of individual triplets, allowing the model to generalize to new TFs and cell types. Across 690 human ChIP-seq experiments covering 161 TFs and 91 cell types, TransBind2 achieves a macro AUROC of 0.9648 and AUPR of 0.4215, outperforming TransBind and other baselines, with a [≥]12.67% relative AUPR gain. The model trained on human data also performs well in cross-species zero-shot prediction on mouse data. Saliency analysis shows that it can identify TF-binding peaks with a median error of 12-38 base pairs (bps) despite being trained on window-level labels. Ablation studies further show that TF structure, chromatin accessibility, and bidirectional attention each improve performance. Overall, these results show that combining TF structure with chromatin context leads to more accurate and generalizable TF-DNA binding predictions.
ER -