Mutational Signature

Mutational signatures and the COSMIC Single-Base Substitution (SBS) catalog provide a mathematical and biochemical framework for decomposing the somatic mutations accumulated in a cell genome into the distinct endogenous and exogenous DNA-damaging and DNA-repair processes that generated them. Throughout life, somatic DNA is subjected to spontaneous chemical reactions (such as hydrolytic deamination of 5-methylcytosine), enzymatic editing (APOBEC cytidine deaminases, somatic hypermutation), environmental mutagens (ultraviolet photons, tobacco polycyclic aromatic hydrocarbons, aristolochic acid), and replication or repair defects (mismatch repair deficiency, homologous recombination deficiency, polymerase proofreading mutations). Each process preferentially damages or misrepairs specific local nucleotide sequences, leaving a characteristic quantitative fingerprint across the genome.

Pioneered by Serena Nik-Zainal, Ludmil B. Alexandrov, Michael R. Stratton, and colleagues (2012–2020), the COSMIC (Catalogue Of Somatic Mutations In Cancer) mutational signature framework represents single-base substitutions as a 96-component probability vector conditioned on the mutated pyrimidine base and its immediate 5′ and 3′ flanking nucleotides, supplemented by 78-channel doublet-base substitution (DBS) and 83-channel small insertion/deletion (ID) classifications. By applying Non-Negative Matrix Factorization (NMF) across tens of thousands of whole-genome (WGS) and whole-exome (WES) sequences and Non-Negative Least Squares (NNLS) refitting to individual tumors, mutational signature analysis quantifies the active mutagenic exposures inside a tissue, guides homologous recombination (PARP inhibitor) and immunotherapy stratification, and separates neutral mutational bias from positive selection and purifying selection in codon substitution models.

Strand Symmetry Reduction from 192 to 96 Trinucleotide Channels

At any genomic base pair, there are 4 possible reference nucleotides (A, C, G, T), each of which can mutate into 3 alternative bases, yielding 4 × 3 = 12 apparent base substitutions. However, in double-stranded B-form DNA prior to strand-specific transcription-coupled analysis, a substitution on the purine strand is Watson-Crick complementary to a substitution on the opposing pyrimidine (C or T) strand (for example, a G>T transversion on the plus strand is identical to a C>A transversion on the complementary minus strand). Standardizing every single-base substitution by the mutated pyrimidine (C or T) collapses the 12 base changes into 6 canonical base substitution classes:

• Three cytosine (C) substitutions: C>A (transversion), C>G (transversion), and C>T (transition).
• Three thymine (T) substitutions: T>A (transversion), T>C (transition), and T>G (transversion).

Because DNA polymerases, glycosylases, and chemical adducts are strongly influenced by base-stacking interactions with the immediately adjacent nucleotides, each of the 6 pyrimidine substitutions is subclassified by the 4 possible 5′ bases (A, C, G, T) and 4 possible 3′ bases (A, C, G, T) on the pyrimidine-containing strand (spanning 4 × 4 = 16 flanking dinucleotide contexts, written in bracket notation such as A[C>T]G or T[C>A]T). Multiplying the 6 substitution classes by the 16 trinucleotide contexts defines the canonical 96-channel SBS mutational spectrum:

m = (m₁, m₂, ..., m₉₆)^T ∈ ℝ_≥0⁹⁶, normalized to a probability simplex ∑_i=1..96 p_i = 1

When comparing mutational spectra captured from Whole-Exome Sequencing (WES, ~2% of the human genome, enriched in GC-rich coding exons) against Whole-Genome Sequencing (WGS, ~41% GC across 3.1 Gb), observed counts m_i are divided by the reference genome or exome 3-mer opportunity frequency f_i^3-mer so that signature profiles reflect intrinsic per-trinucleotide mutation rates rather than genomic base composition.

Information-Theoretic Spectrum Metrics: Shannon Entropy and Jensen-Shannon Divergence

The concentration or diffuseness of a 96-channel signature p = (p₁, ..., p₉₆) is quantified by its 96-channel Shannon entropy (H₉₆) in bits, bounded above by log₂(96) = 6.585 bits for a flat, uniform white-noise spectrum:

H₉₆(p) = − ∑_i=1..96 p_i log₂(p_i) and Normalized Flatness J′ = H₉₆(p) / log₂(96) ∈ [0, 1]

Focal enzymatic or chemical processes that target a single motif—such as SBS1 (spontaneous deamination of 5-methylcytosine at N[C>T]G CpG dinucleotides) or SBS2 (APOBEC deamination at T[C>T]N)—exhibit low entropy (H₉₆ ≈ 2.2 to 3.8 bits, concentrating >75% of mutations into 4 of the 96 channels). In contrast, loss of homologous recombination (SBS3, BRCA1/BRCA2 deficiency) or flat clock-like background mutagenesis (SBS5 and SBS40) exhibits high entropy (H₉₆ ≈ 5.9 to 6.4 bits, distributed broadly across all 96 channels).

Similarity between an observed tumor spectrum p and a reconstructed or reference COSMIC signature q is evaluated using cosine similarity—cos(θ) = (p · q) / (||p||₂ ||q||₂) ∈ [0, 1] (where cos(θ) ≥ 0.90 indicates a high-confidence reconstruction)—and the bounded, symmetric Jensen-Shannon Divergence (JSD) in bits:

JSD(p || q) = (1 / 2) D_KL(p || M) + (1 / 2) D_KL(q || M), where M = (1 / 2)(p + q) and 0 ≤ JSD(p || q) ≤ 1.0 bit

De Novo Signature Discovery via Non-Negative Matrix Factorization (NMF)

Given a cohort of G sequenced tumor genomes, their observed 96-channel mutation counts are assembled into a non-negative 96 × G catalog matrix M, where column g contains the 96 trinucleotide counts of genome g. Because both mutation probabilities and total mutational burdens are physically non-negative, de novo mutational signature extraction (SigProfilerExtractor, Alexandrov et al., 2013) models M as the linear product of a 96 × K signature definition matrix P and a K × G exposure (activity) matrix E:

M ≈ P · E, where P_ik ≥ 0, ∑_i=1..96 P_ik = 1 for each signature k ∈ {1, ..., K}, and E_kg ≥ 0

Here, column p_k of P represents the 96-channel probability profile of the k-th mutational signature, and entry E_kg represents the absolute number of somatic mutations contributed by signature k to genome g. To determine the true number of active biological signatures K and estimate (P, E), SigProfilerExtractor performs Poisson resampling of M across 100–1,000 bootstrap replicates, minimizes the generalized Kullback-Leibler divergence (equivalent to maximizing a Poisson likelihood M_ig ~ Poisson((P E)_ig)) or Frobenius norm via Lee-Seung multiplicative update rules, and clusters the resulting solutions using k-medoids (silhouette width ≥ 0.85 alongside minimal reconstruction error) to select the rank K.

Single-Sample Signature Assignment via Non-Negative Least Squares (NNLS)

Once the reference catalog matrix P (such as the COSMIC v3.4 catalog comprising 86 validated SBS signatures) has been established from large pan-cancer cohorts, analyzing a single new patient tumor (with 96-element count vector m ∈ ℝ_≥0⁹⁶) does not require re-running NMF. Instead, signature refitting solves the convex constrained optimization problem of Lawson-Hanson Non-Negative Least Squares (NNLS) to find the non-negative exposure vector e = (e₁, ..., e_K)^T ≥ 0 that minimizes the squared Euclidean residual:

min_{e ≥ 0} || P e − m ||₂²

Active-set Lawson-Hanson NNLS partitions the K candidate reference signatures into an active passive set P (e_k > 0) and a zero-clamped set Z (e_k = 0), iteratively moving the signature with the largest positive gradient P^T(m − P e) into the active set and solving the unconstrained least-squares subproblem via QR decomposition until all Karush-Kuhn-Tucker (KKT) optimality conditions are satisfied. To prevent overfitting ("signature bleeding," where flat signatures or rare tissue-inappropriate signatures absorb random Poisson noise), modern refitting pipelines (SigProfilerAssignment, deconstructSigs, MutationalPatterns) apply backward-forward stepwise pruning: any signature whose removal reduces cosine similarity by less than Δcos(θ) = 0.01 or contributes <5% of total mutations is pruned to zero.

Mutational signature decomposition has revealed fundamental mechanisms of somatic evolution across both normal human tissues and malignancies:

1. Universal Clock-Like Signatures (SBS1 and SBS5): Across normal human stem cells (colonic crypts, hematopoietic progenitors, neurons, and germline cells), two signatures accumulate linearly with chronological age at tissue-specific annual rates. SBS1 arises from spontaneous hydrolytic deamination of 5-methylcytosine (5mC) at methylated CpG dinucleotides directly into thymine, creating T:G mismatches that—if unrepaired by thymine-DNA glycosylase (TDG) or MBD4 before S-phase DNA replication—fix as C>T transitions at N[C>T]G sites in proportion to cumulative mitotic divisions. SBS5 is a broad, clock-like signature enriched in T>C and C>T substitutions that accumulates linearly with age even in post-mitotic neurons, reflecting continuous endogenous DNA damage and error-prone translesion synthesis.
2. Episodic APOBEC Mutagenesis and Kataegis (SBS2 and SBS13): Members of the AID/APOBEC cytidine deaminase family (primarily APOBEC3A and APOBEC3B, which evolved as innate restriction factors against retroviruses and retrotransposons) deaminate cytosine to uracil strictly within single-stranded DNA at TCN motifs (T[C>U]A and T[C>U]T). Replication directly across the uracil produces T[C>T]N transitions (SBS2), whereas excision of uracil by uracil-DNA glycosylase (UNG) generates an abasic (AP) site that, when bypassed by the translesion cytidine transferase REV1, yields T[C>G]N and T[C>A]N transversions (SBS13). When APOBEC3A/B attacks persistent tracts of single-stranded DNA exposed at stalled replication forks or chromothriptic double-strand break resection intermediates, it generates kataegis ("thunderstorms")—localized hypermutation clusters of dozens of processive, strand-coordinated C>T and C>G mutations within a 1 to 10 kb window.
3. Synthetic Lethality and Immunotherapy Stratification (SBS3, SBS44, SBS10): Because a driver gene can be silenced by promoter hypermethylation or complex structural rearrangements missed by targeted exon panels, mutational signatures provide a functional readout of pathway loss. Detection of SBS3 (accompanied by microhomology-flanked deletions ID6 and homologous recombination deficiency scores such as HRDetect) identifies BRCAness in breast, ovarian, pancreatic, and prostate tumors lacking germline BRCA1/2 mutations, predicting clinical response to PARP inhibitors. Similarly, SBS44/SBS6 (mismatch repair deficiency) and SBS10a/b (POLE proofreading deficiency) generate thousands of neoantigenic frameshift or missense peptides that escape nonsense-mediated decay, rendering these tumors responsive to immune checkpoint blockade (anti-PD-1).

1. Alexandrov, L. B., et al. "Signatures of mutational processes in human cancer." Nature 500, 415–421 (2013). https://doi.org/10.1038/nature12477
2. Alexandrov, L. B., et al. "The repertoire of mutational signatures in human cancer." Nature 578, 94–101 (2020). https://doi.org/10.1038/s41586-020-1943-3
3. Nik-Zainal, S., et al. "Mutational processes molding the genomes of 21 breast cancers." Cell 149, 979–993 (2012). https://doi.org/10.1016/j.cell.2012.04.024
4. Alexandrov, L. B., et al. "Clock-like mutational processes in human somatic cells." Nature Genetics 47, 1402–1407 (2015). https://doi.org/10.1038/ng.3441