Position Weight Matrix (PWM)

The core of a PWM is the log-odds score, which compares the observed frequency of a nucleotide at a position to a background frequency. A positive score indicates a nucleotide is favored for binding, while a negative score indicates it is disfavored. This transforms raw counts into a biophysically meaningful energy metric.

• Formula: Score(b, i) = log2( f(b,i) / p(b) )
• Background Model: Typically a uniform distribution (0.25 for each base) or a genome-wide nucleotide frequency.

A PWM is built from a multiple sequence alignment of known binding sites. The process involves:

• Counting: Tallying the frequency of A, C, G, and T at each column of the alignment.
• Pseudocounts: Adding a small constant to each count to avoid zero probabilities and infinite log-odds scores for unobserved nucleotides.
• Normalization: Dividing by the total number of sequences to get a Position Frequency Matrix (PFM), then converting to log-odds.

The information content (IC) at each position measures the degree of conservation, ranging from 0 bits (no preference) to 2 bits (absolute conservation of a single nucleotide). It is calculated as the difference between maximum possible entropy and the observed entropy.

• Total IC: The sum of IC across all positions indicates the overall specificity of the motif.
• Sequence Logos: A visual representation where the height of each letter is proportional to its IC, making binding preferences immediately apparent.

PWMs are used as scanning windows to predict transcription factor binding sites across a genome. A score is calculated for every possible sub-sequence of length L by summing the log-odds values for the nucleotide at each position.

• Thresholding: A match is called when the cumulative score exceeds a statistically defined threshold, often determined by a p-value based on the score distribution of background sequences.
• Computational Efficiency: The linear-time scanning algorithm makes PWMs suitable for analyzing entire mammalian genomes.

A raw PWM score is not directly comparable between different matrices. The p-value associated with a score quantifies the probability of observing a score that high or higher in a random genomic background. This allows for standardized, false-positive-controlled binding site prediction.

• Score Distribution: The distribution of scores for random sequences can be modeled using dynamic programming or extreme value distributions.
• Multiple Testing Correction: When scanning millions of base pairs, corrections like Bonferroni or Benjamini-Hochberg are essential.

The standard PWM model assumes positional independence, meaning the contribution of a nucleotide at one position does not depend on the identity of nucleotides at other positions. This simplification ignores complex biophysical realities.

• Inter-dependent Positions: Dinucleotide correlations and DNA shape readout are not captured.
• Low-Affinity Sites: PWMs often miss functional, low-affinity binding sites that fall below a strict score threshold.
• Context Effects: Flanking sequence and chromatin accessibility, which heavily influence in vivo binding, are not modeled.

A graphical representation of a Position Weight Matrix where the height of each nucleotide letter is proportional to its information content at that position. Taller stacks indicate higher conservation, while the relative heights of individual letters within a stack show the preference for each base. Sequence logos provide an intuitive, visual summary of a transcription factor's binding specificity, making it easy to compare motifs across different proteins or experimental conditions.

A numerical representation of DNA sequence where each nucleotide (A, C, G, T) is mapped to a binary vector of length four. For example, A = [1,0,0,0], C = [0,1,0,0]. This converts a character string into a sparse matrix suitable for neural network input. A PWM can be viewed as a learned weight matrix applied directly to this one-hot encoded representation, where the dot product yields a binding score.

A specific DNA sequence, typically 6–20 base pairs in length, where a transcription factor protein physically binds to regulate gene expression. A PWM is the mathematical model that captures the binding specificity of a transcription factor, representing the probability distribution of nucleotides across all valid binding sites. The PWM is used to scan a genome and predict novel TFBS locations.

The fundamental statistical unit of a PWM. Each entry is calculated as log₂(P_ij / B_i), where P_ij is the probability of nucleotide i at position j in the binding site model, and B_i is the background frequency of that nucleotide in the genome. A positive score indicates a nucleotide is favored over random chance, while a negative score indicates it is disfavored. The sum of log-odds scores across a sequence window gives the total binding affinity estimate.

A computational perturbation method that systematically introduces virtual nucleotide substitutions into a DNA sequence and measures the resulting change in a model's binding prediction. When applied with a PWM, this quantifies the effect of single nucleotide variants on binding affinity. A mutation that drastically reduces the PWM score is predicted to be a functional regulatory variant that disrupts transcription factor binding.