TransBind

Transcription factors bind DNA in a sequence-specific manner, and mapping where they bind across a genome is central to understanding gene regulation. Almost every deep learning method built for this task — DeepSEA, DanQ, TBiNet — frames it as multi-label classification over a fixed panel of ChIP-seq experiments: a DNA window goes in, and one output unit per TF–cell type pair comes out. The identity of the transcription factor enters the model only as an output index, never as an input. Two limitations follow. The model cannot exploit anything known about the protein's sequence or DNA-binding domain, and it can say nothing at all about a transcription factor absent from the training panel.

TransBind, from Shreya Basnet and Jianlin Cheng at the University of Missouri, makes the transcription factor a first-class input. Each TF is represented by an embedding from ESM-DBP, a protein language model created by domain-adaptive pretraining of ESM-2 on roughly 170,000 DNA-binding protein sequences. A cross-attention module lets that protein embedding query the genomic representation directly, so each factor attends to the regions its own binding properties favor. The payoff is twofold: higher accuracy on the standard 690-experiment ENCODE benchmark, and a model that can be pointed at a transcription factor it was never trained on.

The work was posted as a preprint in September 2025 and published in NAR Genomics and Bioinformatics in 2026. Source code and trained checkpoints are released under GPL-3.0.

#Key Features

• Protein-conditioned prediction: The transcription factor enters as a learned representation of its amino acid sequence rather than a one-hot label, so the model can relate binding preferences across structurally similar factors.
• Cross-attention fusion: The TF embedding acts as the query against position-wise DNA features as keys and values, producing a TF-specific weighting of the input window instead of a single shared representation for all factors.
• Zero-shot prediction for unseen TFs: A companion binary-classification variant, trained with TF-embedding masking and Gaussian noise injection, accepts an arbitrary TF–DNA pair and predicts binding without any retraining or fine-tuning.
• Motif-level interpretability: 160 of the 320 first-layer convolutional kernels match known JASPAR motifs, and TF-conditioned attribution recovers cognate motifs across zinc-finger, bZIP, bHLH, E2F, Forkhead, and STAT families.
• Compact and released: At roughly 4.6M parameters the model is small enough to run on a single GPU, and both the multi-label and zero-shot checkpoints ship in the repository.

#Technical Details

The DNA arm takes a one-hot encoded 1,000 bp window — a 200 bp bin extended by 400 bp on each flank — through a 1D convolution with 320 filters of width 26, max-pools with a window of 13 down to 75 positions, then applies a two-layer bidirectional LSTM with 160 hidden units per direction and a 16-head transformer block. The protein arm projects the 1,280-dimensional ESM-DBP embedding into the same 320-dimensional space, and 16-head cross-attention combines the two.

Training uses the ENCODE-derived DeepSEA benchmark on GRCh37: 690 ChIP-seq experiments covering 161 TFs and 91 human cell types, split by chromosome (chr1–6, 10–22 and X for training; chr7 for validation; chr8–9 for test) with reverse-complement augmentation to 3.8M samples. Positive bins are sparse, about 1.44%. TransBind reaches AUROC 0.9508 and AUPR 0.3741, against 0.9402/0.3346 for TBiNet, 0.9490/0.3260 for EPBDxDNABERT-2 (which augments DNABERT-2 with DNA breathing dynamics), 0.9254/0.3065 for DanQ, and 0.8934/0.2509 for DeepSEA — a relative AUPR gain of at least 11.8%, and a win in 676 of 690 TF–cell type pairs on AUPR.

#Applications

The multi-label model serves the usual regulatory-genomics workflow: scoring candidate enhancers and promoters, annotating the likely factor occupancy of a region, and prioritizing non-coding variants by the change they induce in predicted binding. The zero-shot variant addresses a case the panel-based models cannot reach at all — a transcription factor with no ChIP-seq data, including poorly characterized human factors and engineered or designed DNA-binding proteins, where the only inputs available are an amino acid sequence and a genomic window. That makes it useful for triaging which factors merit an expensive ChIP-seq experiment.

#Impact

TransBind's contribution is architectural rather than one of scale: it demonstrates that conditioning a binding-site predictor on a protein language model embedding both improves accuracy on the well-worn 690-experiment benchmark and unlocks a capability the multi-label formulation structurally forbids. The limitations are stated plainly by the authors. Zero-shot accuracy is modest — roughly 0.72 macro AUROC under leave-three-out cross-validation and 0.575–0.629 on three fully held-out factors — and they characterize it as label-zero-shot rather than sequence-zero-shot, since the underlying protein language model may have encountered the held-out sequences during pretraining. The model uses no chromatin accessibility, 3D genome contacts, or predicted TF structures, the zero-shot formulation carries no cell-type information, and training is human-only, leaving cross-species transfer untested.