Transcription factor–DNA binding predictor fusing sequence, DNase-seq accessibility and structure-aware TF embeddings with bidirectional attention.
No providers recorded yet. Browse all providers
The same transcription factor binds different sites in a liver cell and a lymphocyte,
because most of its candidate motifs sit in chromatin that one cell type has packed away.
A predictor that sees only DNA sequence has to average over those contexts, and a predictor
that treats each factor as a fixed output column cannot say anything about a factor it was
not trained on. TransBind2 addresses both problems at once: it asks, for a single
<DNA bin, TF, cell type> triplet, whether that factor is bound there, with chromatin
accessibility supplying the cell context and a protein language model supplying the factor.
It was developed by Shreya Basnet and Jianlin Cheng at the University of Missouri as the successor to TransBind, and posted to bioRxiv in September 2026. TransBind had already made the factor an input rather than an output index, but it reduced each factor to one mean-pooled ESM-2-derived vector, let only DNA attend to protein, and needed a separately trained variant for unseen factors. TransBind2 makes four changes: it adds DNase-seq and mappability tracks, switches to per-residue ProstT5 embeddings that carry structure, makes the attention bidirectional, and recasts the task as binary triplet classification so that new factors are handled by the one model.
The DNA encoder passes a 1,000 bp six-channel window through a 320-filter convolution (kernel 26), max pooling, a two-layer bidirectional LSTM and a 16-head transformer layer, then interpolates to 256 positions of width 320. The protein encoder projects ProstT5 features down to 320 dimensions and compresses them with 160 learned query vectors into a fixed 160×320 block, independent of protein length. After two bidirectional cross-attention layers with 8 heads, the DNA stream is fused residually with the original encoding, pooled with learned position weights and passed through a 320→512→256→1 head. Training used balanced batches of 3,500, AdamW with cosine annealing, and Optuna hyperparameter search.
The data are the ENCODE-derived set used for TransBind on GRCh37: 690 ChIP-seq experiments covering 161 factors and 91 human cell types, split by chromosome with chr8–9 held out. As triplets this expands to 2.63 billion examples, about 1.4% positive. DNase-seq tracks exist for 64 of the 91 cell types; the other 27 receive a fixed default signal. On the held-out chromosomes TransBind2 reaches macro AUROC 0.9648 and AUPR 0.4215, against 0.9508/0.3741 for TransBind, 0.9264/0.3291 for TFBindFormer, 0.9180/0.2960 for fine-tuned DNABERT-2, and 0.8934/0.2509 for DeepSEA. It beats TransBind on AUPR for 146 of 161 factors. Ablations rank the inputs clearly: dropping the protein branch cuts AUPR to 0.1537, dropping DNase-seq to 0.3660, and dropping mappability to 0.4078.
TransBind2 suits genome-wide occupancy maps for factors and cell types that lack ChIP-seq, prioritisation of regulatory regions for follow-up, and localisation of the likely binding position inside a window: input-times-gradient saliency places peaks a median 12–38 bp from ChIP-seq summits across eight factors from six structural families, against 55–90 bp for a centre-of-window guess. To predict for a new factor, a user needs its UniProt sequence and an AlphaFold structure to generate ProstT5 features; to predict for a new cell type, a DNase-seq track for that cell type. The prediction script accepts a missing DNase track, but then substitutes the fixed no-coverage value, which removes the input that carries cell-type specificity.
The human-trained model, with weights unchanged, was applied to 26 mouse ChIP-seq experiments in CH12 and MEL cells using mouse DNase-seq, mm9 mappability and mouse factor structures, reaching macro AUROC 0.725 and AUPR 0.221. The authors call this label-zero-shot: no mouse binding labels were seen, although tested factors such as Sin3a and E2f4 have human orthologs in the training set, so the test measures transfer to a new genome and chromatin landscape more than to unfamiliar proteins. Mouse accuracy varies widely by factor, with AUROC from 0.52 to 0.87. The work is a single preprint that has not been peer reviewed, and all evaluation is computational against existing ChIP-seq peaks.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.