TFClassPredict

Reading the cis-regulatory code — deciding which transcription factor (TF) binds where along the genome — is a foundational problem in genomics, and most computational methods frame it as a binary question: is this locus bound or unbound by a given TF? That framing sidesteps a biological reality that makes single-TF assignment inherently ambiguous. Hundreds of human TFs share the same DNA-binding domain (DBD), and factors within a DBD family recognize near-identical sequence motifs, so a prediction that a specific TF binds a site is rarely separable from the many paralogs that would bind it equally well.

TFClassPredict, developed at the University Medical Center Göttingen and released as a December 2024 bioRxiv preprint, reframes the task to match this structure. Instead of predicting individual TFs, it casts binding-site prediction as a multi-class problem over the 23 DBD-classes defined by the TFClass hierarchy, discriminating each class against all others. By working at the level where sequence preference is actually determined — the DNA-binding domain — the model resolves the competitive-binding ambiguity that per-TF predictors inherit, while still pinpointing which family of factors is responsible for a given regulatory element.

The model builds on DNABERT, a BERT-style language model pretrained on the human genome, fine-tuning it into a fixed classifier that is applied genome-wide without any per-input retraining. It joins a small set of catalog models attacking TF-DNA specificity from different angles, including tf-SFM and TransBind, but is distinguished by its DBD-class formulation and its packaged, ready-to-run distribution.

#Key Features

• DBD-class formulation: Prediction is posed as one-vs-rest classification across 23 DNA-binding-domain classes from the TFClass taxonomy, so outputs map onto biologically defined factor families rather than forcing an arbitrary choice among interchangeable paralogs.
• DBD-DNA co-evolution signal: Training on high-confidence, directly bound sites lets the model exploit the co-evolutionary relationship between binding-domain structure and target DNA, the same regularity that makes single-TF calls ambiguous.
• Fixed, transferable checkpoint: A single fine-tuned DNABERT checkpoint is applied to arbitrary genomic loci; the authors demonstrate transfer beyond the training objective to 3D genome interaction prediction and to recovering lineage-specifying TF programs from immune-cell ATAC-seq data.
• Packaged distribution: The method ships as the installable tfclass_predict PyPI package with a command-line tool and a PredictionManager Python API that score BED-format peak files, plus a precomputed genome-wide prediction set so most users need no GPU.

#Technical Details

TFClassPredict fine-tunes a pretrained DNABERT transformer backbone (built on TensorFlow 2.13) into a 23-way DBD-class classifier. Training uses high-confidence, directly bound TFBSs drawn from curated ChIP-seq-derived sites (a UniBind-derived TFBS set is distributed alongside the weights). At inference the tool accepts a BED file of ATAC-seq or other NGS peaks and returns per-peak or per-window DBD-class binding scores, optionally standardized into a TF-binding-potential-density score. The authors report that it delivers robust DBD-class-level classification, outperforming position-weight-matrix baselines and benchmarked deep-learning architectures, with predictions mapping cleanly onto the biologically defined DBD-classes. Pretrained weights and a precomputed hg38 genome-wide prediction archive are hosted on Zenodo (roughly 0.7 GB for the model and 13.5 GB for the compressed predictions), and precomputed scores let the pipeline run on CPU-only systems.

#Applications

The tool is aimed at regulatory-genomics workflows that start from open-chromatin or other peak data and need to know which factor families act at each region. Feeding ATAC-seq peaks through TFClassPredict yields genome-wide DBD-class annotations that recapitulate known lineage-specifying TF programs in immune cells, and the same scores improve conventional PWM-based scanning pipelines and help predict three-dimensional genome interactions. Because it distributes a fixed model plus precomputed hg38 predictions, wet-lab and computational biologists can annotate their own experiments without training a model or supplying a GPU.

#Impact

TFClassPredict offers a concrete answer to a longstanding awkwardness in TFBS prediction: that per-TF calls promise a specificity the underlying sequence signal cannot support. By moving the prediction target to the DNA-binding-domain level, it turns an ill-posed assignment into a well-defined classification with directly interpretable outputs, and its demonstrated transfer to 3D-genome and ATAC-seq tasks suggests the learned representation captures regulatory grammar beyond its training objective. The work remains a preprint that has not completed peer review, and its scope is limited to the human genome and the 23 TFClass DBD-classes, but the open, packaged distribution — code, weights, and precomputed predictions — lowers the barrier to adoption for regulatory-genomics groups.