Decomposing Cancer Genomes with Non-Negative Matrix Factorization

26 concepts26 papers1 proposal

Biological systems are inherently additive mixtures. When a geneticist sequences a tumor, the resulting data does not reflect a single, isolated biological event. Instead, it captures the combined footprint of multiple overlapping processes: the patient’s chronological aging, historical exposure to environmental carcinogens, metabolic shifts, and the infiltration of diverse immune cell types.

Unmixing this signal is one of the central challenges of computational oncology. To solve it, researchers rely heavily on Non-negative Matrix Factorization (NMF). NMF provides an unsupervised framework that naturally mirrors the additive reality of biology, allowing scientists to isolate distinct mutational signatures and molecular subtypes from high-dimensional genomic data.


The Intuition: Why Non-Negativity Matters

Mathematically, we can represent our observed data as a large matrix $V$, where rows correspond to biological features (such as mutation types or gene expression levels) and columns represent individual patient samples. Our goal is to factorize this matrix into two smaller, latent matrices:

$$V \approx W \times H$$

  • $W$ (The Basis Matrix): Represents the fundamental biological patterns, such as distinct mutational signatures or cellular "meta-programs."
  • $H$ (The Coefficient Matrix): Represents the "exposure" or weight of each signature within each individual sample.
   Observed Data (V)            Basis Vectors (W)         Coefficients (H)
  (Features x Samples)        (Features x Signatures)   (Signatures x Samples)

     [ .  .  .  . ]                 [ .  . ]                 [ .  .  .  . ]
     [ .  .  .  . ]      \approx     [ .  . ]        \times   [ .  .  .  . ]
     [ .  .  .  . ]                 [ .  . ]

Standard dimensionality reduction techniques like Principal Component Analysis (PCA) solve this by finding orthogonal axes of variation, which frequently introduces negative values into both $W$ and $H$. However, a negative mutation count or a negative gene expression level makes no physical sense.

NMF enforces a strict constraint: all elements in $W$ and $H$ must be greater than or equal to zero ($W \ge 0, H \ge 0$). Because subtraction is forbidden, the algorithm can only reconstruct the observed data by adding positive components together. This non-negativity constraint eliminates cancellation effects and forces the model to learn a parts-based, highly interpretable representation. The resulting factors map directly onto real physical entities, such as the specific DNA damage caused by ultraviolet light or the transcriptomic profile of an infiltrating T-cell.


Unmixing the History of DNA Damage

Every agent that damages DNA leaves a characteristic scar. For instance, tobacco smoke causes specific base substitutions, while defective mismatch repair mechanisms leave a different trail of genomic errors. NMF allows researchers to take a chaotic catalog of mutations from hundreds of tumors and isolate these distinct chemical and enzymatic histories.

To streamline this process, Hamid R. Rabiee et al. developed CANCERSIGN, an automated tool that applies NMF across varying nucleotide contexts (such as 3-mer and 5-mer configurations) to extract latent biological processes directly from raw mutation matrices.

Once extracted, these signatures provide a direct bridge between environmental exposures and genomic instability:

  • Environmental Toxins: In oesophageal squamous-cell carcinoma, Jiang Chang et al. utilized NMF to isolate a specific mutational fingerprint, termed Signature E4, which correlated tightly with alcohol consumption and genetic variants in alcohol-metabolizing enzymes. Similarly, Xiangchun Li et al. applied NMF to liver cancer genomes to map lifestyle-driven DNA damage patterns back to their underlying molecular mechanisms.
  • Endogenous Aberrations: NMF can also capture internal cellular machinery gone awry. Daniel Hübschmann et al. used the factorization method to show how the natural antibody-diversification machinery in B-cells can cause collateral genomic damage, isolating distinct signatures for class-switch recombination errors in lymphomas.
  • Microenvironmental Stressors: The physical state of the tumor tissue itself alters the genome. Keun Soo Ahn demonstrated that tumor hypoxia (low oxygen) acts as a universal driver of genomic instability, using NMF to identify distinct mutational fingerprints that separate hypoxic tumors from normoxic ones across 27 cancer types.

By decomposing these signals, researchers frequently discover entirely new biology. For example, Kyle M. Hernandez et al. applied NMF to advanced prostate cancer data within the Veterans Precision Oncology Data Commons, uncovering de novo mutational signatures unique to specific cohorts that were completely absent from standard reference databases.


Mapping Cellular States and Subtypes

Beyond DNA damage, NMF is widely used to decompose transcriptomic and proteomic matrices. In this context, the basis vectors in $W$ represent co-expressed gene modules—often called meta-programs—that reflect specific functional states, while $H$ groups patients based on how strongly they express those programs.

Haoxiang Wu et al. demonstrated this utility by applying NMF to single-cell transcriptomic data from small cell carcinoma of the esophagus, distilling highly heterogeneous malignant cell populations into discrete, conserved functional modules. Similarly, in cervical cancer, Zhen Zhang et al. leveraged NMF to cluster complex expression data into meta-programs that define the functional states of the surrounding tumor immune microenvironment.

When these meta-programs are used to stratify patients, they reveal stable, clinically relevant molecular subtypes that traditional clinical staging often misses:

Cancer Type Stratification Basis Key Findings / Clinical Insights
Ovarian Cancer Transcriptomics & Proteomics Hyondeog Kim et al. identified four distinct transcriptomic subtypes, highlighting an "immunoreactive" group poised for immunotherapy. At the proteomic level, Rong Tan et al. used NMF to reduce complexity and isolate subtype-specific therapeutic vulnerabilities like CD40 and TYMP.
Digestive Cancers MicroRNA Expression Senwei Tang et al. decomposed miRNA expression matrices into additive components, creating a stable "molecular compass" that predicts survival outcomes across nearly 1,800 samples.
Liver Cancer Metabolic & Infectious Profiles Yue Li et al. used NMF to group hepatocellular carcinoma (HCC) patients based on amino acid metabolism genes. For parasitic-driven malignancies, Shuyan Sheng et al. applied NMF to identify stable patient subgroups linked to Schistosoma japonicum infections.
Bladder Cancer & Sarcoma Immune-Metabolic Pathways Shao-Yu Yue et al. performed unsupervised clustering on combined immune and metabolic gene sets to stratify bladder cancer risks. For sarcomas, Yuheng Hong et al. decomposed lipid metabolism profiles, isolating a distinct subgroup characterized by high SQLE expression.
Gliomas Cellular Transition Programs Jiasheng Wu et al. applied NMF to gene expression data to cluster lower grade gliomas based on epithelial-mesenchymal transition (EMT) patterns, revealing how surrounding normal tissue influences tumor malignancy.

Generalizing Beyond Bulk Sequencing

Because NMF is a general mathematical framework for additive decomposition, its utility extends far beyond traditional bulk sequencing. It scales seamlessly to spatial and phenotypic modalities that capture the physical architecture of tissues.

In single-cell spatial imaging, for instance, the challenge shifts from unmixing molecules to unmixing cellular neighborhoods. Bing He et al. applied NMF to spatial interaction features extracted from breast cancer tumors. By decomposing these spatial "social networks" of cells, their framework identified seven distinct survival subtypes and successfully isolated atypical, high-risk patients who unexpectedly demonstrated positive clinical outcomes.

  Spatial Tissue Grid         NMF Decomposition          Latent Spatial Signatures
  [Tumor][Immune][Stromal]                           [Neighborhood A: Immune-Rich]
  [Stromal][Tumor][Immune]   ====================>  
  [Immune][Immune][Tumor]                            [Neighborhood B: Vascular Niche]

Similarly, NMF resolves the spatial complexity inherent to metabolomics. Mass Spectrometry Imaging (MSI) generates massive datasets where every pixel contains a dense metabolic spectrum. As reviewed by Ying Cai et al., NMF serves as a critical unsupervised dimensionality reduction tool in this space. By decomposing dense spatial metabolic matrices without prior labels, NMF isolates localized biomarkers and maps distinct metabolic signatures directly to precise anatomical microregions within complex tissues.

Ultimately, NMF works because it respects the physical constraints of biology. By enforcing non-negativity, it translates high-dimensional genomic, proteomic, and spatial mixtures into an interpretable vocabulary of distinct biological parts.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_non_negative_matrix_factorization_mutati_b49f0cec,
  title = {non-negative matrix factorization (mutational signature decomposition)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-negative-matrix-factorization-mutational-signature-decomposition/rdr_b49f0ceca55947e3b1eaa9174acb0074},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_non_negative_matrix_factorization_mutati_b49f0cec_v1,
  title = {non-negative matrix factorization (mutational signature decomposition)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-negative-matrix-factorization-mutational-signature-decomposition/rdr_b49f0ceca55947e3b1eaa9174acb0074/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake