NyxBind

Transcription factor binding sites (TFBS) are short DNA motifs where regulatory proteins dock to control gene expression, and identifying them accurately from sequence is a long-standing problem in regulatory genomics. Pretrained genomic language models such as DNABERT-2 learn general DNA sequence patterns through masked language modeling, but their embeddings often fail to discriminate the subtle, biologically critical differences that separate one transcription factor's motif from another. NyxBind addresses this gap by reshaping a genome foundation model's representation space specifically for binding-site discrimination.

Developed by researchers at the Hong Kong University of Science and Technology (Guangzhou) and released as a preprint in October 2025, NyxBind is the first TFBS prediction model to apply contrastive learning across diverse TFBS types. Rather than training a classifier directly, it continues pretraining DNABERT-2 with a contrastive objective that pulls together sequences bound by the same factor while pushing dissimilar sites apart, yielding representations that are sharper for downstream binding-site tasks.

The result is a reusable pretrained backbone that can be fine-tuned—fully or with parameter-efficient methods—for individual transcription factors, while also supporting interpretable motif visualization.

#Key Features

• Contrastive representation learning: NyxBind refines DNABERT-2 embeddings with contrastive pretraining over sequence pairs spanning 160 TFBS types, sharpening its ability to distinguish closely related motifs.
• Foundation-model backbone: Built on the cataloged DNABERT-2 (117M) genome language model, it inherits general DNA sequence knowledge while specializing the representation space for regulatory sequences.
• Flexible fine-tuning: Supports both full-parameter fine-tuning and LoRA-based parameter-efficient adaptation, making per-factor models inexpensive to train.
• Motif interpretability: Attention-based visualization recovers reliable sequence motifs, connecting predictions back to recognizable biological signals.

#Technical Details

NyxBind builds on the DNABERT-2 transformer (117M parameters) with a customized BERT layer implementation for contrastive pretraining. Contrastive training uses sequence pairs drawn from 160 TFBS types, with chromosomes 1–23 (excluding 11 and 12) plus X and Y used for training and chromosomes 11–12 held out for validation and testing. Task-specific fine-tuning is evaluated on 33 ChIP-seq datasets organized by transcription factor. Across these benchmarks the authors report improved ROC-AUC, PR-AUC, accuracy, precision, recall, and F1 relative to CNN-based predictors (DeepBind, DanQ), BERT-TFBS variants, and Nucleotide Transformer baselines. Pretrained weights are distributed through Hugging Face (CompBioDSA/NyxBind).

#Applications

NyxBind targets computational and regulatory genomics workflows where researchers need accurate, transferable TFBS predictions—annotating candidate binding sites genome-wide, prioritizing regulatory variants, and studying how motif grammar governs transcription factor occupancy. Its parameter-efficient fine-tuning and released checkpoints make it practical to adapt to new transcription factors or ChIP-seq datasets without retraining a genome model from scratch, and its motif visualization supports hypothesis generation about the sequence determinants of binding.

#Impact

By demonstrating that contrastive learning across many TFBS types measurably improves a genome language model's binding-site representations, NyxBind offers a template for specializing general-purpose DNA foundation models toward fine-grained regulatory tasks. It extends the DNABERT-2 lineage into contrastive regulatory-genomics territory and provides an openly available pretrained backbone and benchmarking suite. As a preprint awaiting peer review, its improvements are so far established in in-silico benchmarks, and the repository does not currently specify a software license.