Transcription Factor Binding Site (TFBS)

TFBSs are typically 6–20 base pairs in length. Unlike strict digital codes, they exhibit degeneracy, meaning the transcription factor tolerates nucleotide variations at certain positions. This binding flexibility is quantitatively captured by a Position Weight Matrix (PWM) , which assigns a log-odds score to each base at each position, representing the statistical preference derived from aligned binding sites.

A TFBS is functionally defined not just by its sequence, but by its chromatin context. A perfect consensus sequence will not be bound if it is occluded by a nucleosome. Active binding sites are found in open chromatin regions characterized by:

• DNase I hypersensitivity
• ATAC-seq signal These assays identify nucleosome-depleted regions where the DNA backbone is physically accessible to transcription factors.

Transcription factor binding is not a static lock-and-key event. It is a highly dynamic, transient interaction governed by:

• Residence Time: The duration a protein remains bound (milliseconds to minutes).
• On/Off Rates: The kinetic constants of association and dissociation. High-resolution techniques like ChIP-nexus and single-molecule imaging reveal rapid turnover, where factors constantly sample the genome until they find a functional cognate site or are stabilized by co-factors.

Regulatory logic often requires more than one factor. Homotypic clusters (multiple copies of the same motif) and heterotypic clusters (different factors binding adjacently) create a combinatorial grammar. This cooperative binding increases affinity and specificity. Deep learning models like BPNet excel at learning this grammar by using dilated convolutions to integrate signals across these multi-motif enhancer modules.

A single nucleotide variant within a TFBS can drastically alter binding affinity. Allele-Specific Binding (ASB) occurs when a heterozygous variant causes differential binding between maternal and paternal alleles. This provides direct functional evidence linking non-coding genetic variation to disease risk. In silico mutagenesis systematically predicts these effects by introducing virtual substitutions into a trained neural network and measuring the predicted binding change.

While the DNA double helix is anti-parallel, the binding site itself is asymmetric. A transcription factor recognizes a specific 5' to 3' orientation. Computational models must account for this by scanning both strands. The reverse complement of a motif is a distinct feature, and data augmentation techniques that randomly reverse-complement sequences are critical for enforcing strand symmetry in genomic deep learning models.

Chromatin Immunoprecipitation followed by sequencing, an experimental assay that combines antibody-based enrichment of protein-bound DNA with high-throughput sequencing to map protein-DNA interactions genome-wide.

• Antibodies target specific transcription factors or histone modifications
• Enriched DNA fragments are sequenced and aligned to a reference genome
• Produces signal peaks at regions of protein occupancy
• Requires downstream peak calling algorithms to distinguish signal from background noise

A pioneering deep convolutional neural network architecture that predicts sequence specificities of DNA- and RNA-binding proteins directly from raw nucleotide sequences. DeepBind uses multiple parallel convolutional filters to learn binding motifs without prior knowledge.

• Accepts one-hot encoded sequences as input
• Each convolutional filter learns a distinct sequence motif
• Outputs a binding probability score for any input sequence
• Demonstrated that deep learning could outperform traditional PWM-based methods

The unsupervised computational process of identifying recurring, statistically overrepresented sequence patterns from a set of unaligned DNA sequences. Unlike PWM scanning, motif discovery requires no prior knowledge of the binding preference.

• Uses expectation-maximization or Gibbs sampling algorithms
• Iteratively refines motif models and binding site locations
• Outputs a position weight matrix and set of predicted binding sites
• Essential for characterizing novel transcription factors with unknown specificity

A computational perturbation method that systematically introduces virtual nucleotide substitutions into a DNA sequence and measures the resulting change in a neural network's binding prediction. This technique identifies causal regulatory variants without wet-lab experiments.

• Every position is mutated to all three alternative nucleotides
• The predicted binding change quantifies each position's importance
• Reveals allele-specific binding effects of single nucleotide variants
• Complements experimental methods like MPRA for variant interpretation

The computational process of analyzing ChIP-seq or ATAC-seq signal profiles to identify genomic regions with statistically significant enrichment of mapped reads over background noise. Peak callers distinguish true binding events from experimental artifacts.

• Models background using control input DNA or Poisson distributions
• Accounts for GC bias and mappability artifacts
• Outputs ranked peak lists with significance scores (p-values, q-values)
• IDR framework assesses reproducibility across biological replicates