Aksu TFBS General Model

The human genome encodes roughly 1,600 transcription factors, and mapping where any one binds requires a ChIP-seq experiment with an antibody against that protein. Most factors have no such experiment in most cell types, and the models built to fill the gap inherit the limitation: fit to one TF's ChIP-seq labels, they are undefined for a factor that has no labels at all.

Ekin Deniz Aksu and Martin Vingron at the Max Planck Institute for Molecular Genetics invert the usual training layout. Rather than fitting one model per factor, they pool every TF–cell type pair in the training set into a single TF-agnostic "general model" whose only factor-specific inputs are the TF's binding motif and its mRNA expression — two quantities available for essentially any human TF. The resulting fixed checkpoint scores factors, cell types and chromosomes all absent from training, without being re-fit. Its output then feeds back as an input feature when fine-tuning per-TF models, making the general model both a standalone predictor and the transfer-learning vehicle for the specialized ones.

The work carries no product name in the preprint or the peer-reviewed version; the paper title and the tfbs_prediction_paper repository are its identifying handles. It sits alongside DNA language models such as Nucleotide Transformer and DNABERT rather than competing with them: Nucleotide Transformer supplies the frozen embeddings the per-TF variants consume as ordinary features.

#Key Features

• TF-agnostic general model: One fixed checkpoint predicts binding for factors it never saw in training, given only their motif and expression.
• Two-stage transfer learning: The general model's score enters the TF-tuned models as an input feature; a fitted tree short-circuits on confident scores and probes accessibility and co-binding only in the uncertain middle range.
• Frozen DNA language model embeddings: The TF+transformer variant appends 1,024 Nucleotide Transformer dimensions to the handcrafted block, lifting accuracy, though the embeddings alone fall well short of the full feature set.
• Biologically legible features: 34 handcrafted features span genomic (conservation, ReMap signal), motif (maximum PWM score, TRAP total affinity), cell-type (ATAC-seq, CRUP enhancer probability) and cell-and-TF families (TF expression and activity, TOBIAS footprinting, co-binding).
• Deliberately narrow context: Scoring happens per candidate cis-regulatory element, at most 350 bp, trading the multi-kilobase receptive field of DREAM-era convolutional models for interpretability.

#Technical Details

The head is XGBoost gradient-boosted trees with a logistic objective, 500 estimators, depth 6, learning rate 0.05 and 0.8 row and column subsampling. Training covers roughly 1.5 million hg38 regions — ENCODE candidate cis-regulatory elements plus ReMap peaks and random background — across 8 cell types, 29 TFs and 75 TF–cell type pairs, with chromosomes 1, 8 and 20 held out. On the ENCODE-DREAM final round the TF+transformer model reaches a mean AUPR of 0.505, ahead of J-Team at 0.48 and FactorNet and autosome.ru at 0.46, against 0.02 for random guessing. On ten factors absent from training — NR2F2, ZBTB33, RXRA, RAD21 and HNF4G in liver, ISL1, TFAP2B, FOSL2, NFIC and PBX3 in SK-N-SH — the general model averages 0.36 AUPR and reaches 0.65, beating the median pretrained TF+transformer model on 8 of the 10.

Both the preprint and the journal version place the Nucleotide Transformer embeddings in the per-TF TF+transformer models alone; the preprint attributes that to memory limits when pooling embeddings across all TF–cell type pairs, an explanation the journal version drops. The released code takes the other route: 5_general_model_training.py fits the general model on embeddings and handcrafted features together and writes general_model_nt.json, and the downstream script loads that same checkpoint, retrying the prediction on the handcrafted features alone when the embedding columns are absent.

#Applications

The immediate use is estimating binding where no experiment exists. Given a PWM and an expression measurement in the tissue of interest, the general model returns per-enhancer binding probabilities genome-wide, the input gene-regulatory-network construction otherwise has to guess at. In the paper's liver example, four enhancers in intron 2 of the LEPR gene receive HNF4A probabilities of 0.01, 0.88, 0.81 and 0.30, correctly separating the bound sites from the unbound one. Because predictions decompose into named features, a hit traces back to accessibility, footprinting or motif affinity.

#Impact

The gain from the DNA language model is real but modest, around 0.01 AUPR over the equivalent model without embeddings, and the general model generally trails a freshly trained TF-specific model on the held-out factors, as well as mean ChIP-seq fold-change signal. Ground-truth quality also bounds the result: ChIP-seq sets with a fraction of reads in peaks below 0.02 yield general-model AUPRs under 0.15, and the authors recommend discarding them. The repository is a reproduction pipeline, not an inference package, so scoring new regions means assembling the external feature stack — TRAP, MOODS, TOBIAS, bwtool, bedtools — alongside precomputed CRUP, ReMap and phastCons tracks. What the work establishes is the layout: pooling across factors, then fine-tuning, is a workable route to binding estimates for TFs that will never get their own ChIP-seq experiment.