Research Radar — 2026-10-13

Methodology Review: Nonnegative matrix factorization: learning reusable expression programs

English · About 10 minutes

Primary-source full texts inspected for all readings. The worked example is illustrative; published biological benchmarks were not independently reproduced. The proposed protein-abundance application is explicitly speculative.

The intuition

Imagine a hypothetical expression experiment in which some cells carry an identity-associated pattern and, independently, an activity-associated pattern. A cell can show both. Before naming a new cell type for every combination, it is worth asking a different question: could a small collection of reusable patterns reconstruct the observations? This is a modeling thought experiment. It does not assume that any particular biological dataset has a clean decomposition, or that every recovered pattern deserves a mechanistic name.

Nonnegative matrix factorization, abbreviated NMF, asks for such an additive description. Each pattern has nonnegative feature weights, and each observation receives nonnegative amounts of the patterns. Contributions accumulate without cancellation. Geometrically, combinations of the pattern vectors fill a cone: starting at zero, one can move outward by adding positive amounts of its generating directions. Finding a cone in which the observations can be represented exactly or approximately is easier to state than proving its directions are the true underlying parts. [R4]

A small worked example

Build a tiny example with three measured features and four observations. Let the two columns of \(W\) be invented patterns A and B, and let the columns of \(H\) specify their amounts in observations one through four. The entries of \(W\) and \(X\) use arbitrary signal units; \(H\) contains dimensionless usage coefficients, and the example has no sampling noise. The construction is deliberately chosen to be exactly reconstructible:

Multiplying the two matrices gives the observation matrix \(X\). Rows retain the feature order, while columns retain the observation order:

The first observation contains one unit of A; the second contains two units of A. The third contains one unit of B. The fourth contains one unit of each. For example, the fourth column is obtained by adding the two pattern columns, producing the vector with entries two, one and three.

Notice two distinct kinds of variation. The first and second observations have the same pattern at different amplitudes. The first and fourth share pattern A, but the fourth also contains B. A single label for each observation would omit some of this structure unless the label were supplemented with other measurements. Here the factorization makes both relationships explicit. We supplied the factors by hand; recovering them from the observations alone is a separate problem.

Now multiply every entry of pattern A by ten and divide every amount of A by ten. The reconstructed observations remain unchanged. In observation four, the usage coefficients become 0.1 and 1, even though each pattern still contributes three measurement units. Treating those two coefficients as percentages would give a misleading impression. Factor normalization is part of the interpretation. For this toy example, dividing each pattern by its sum and multiplying its usage row by that sum would put the two contributions on a comparable total-signal scale.

Try two small checks before continuing. A single pattern cannot exactly produce both observation one and observation three, because those vectors are not proportional. Also, change only the first entry of the observed matrix from two to three while keeping the displayed factors fixed. There is then one residual of size one and all other residuals are zero. The half-squared-error objective introduced below becomes one half. This checks the meaning of a residual without confusing fitting the model with evaluating an already chosen fit.

The necessary mathematics

Let the nonnegative matrix \(X\) have \(G\) feature rows and \(N\) observation columns. Choose \(K\) patterns; \(W\) has dimensions \(G\times K\) and \(H\) has dimensions \(K\times N\). Indices \(g,n,k\) identify a feature, observation and pattern. One standard objective is

The inequalities apply entry by entry. The inner sum reconstructs one measurement; the outer sums accumulate squared residuals. Holding either factor fixed gives a convex subproblem, but optimizing both together is nonconvex. [R1]

Lee and Seung give a multiplicative update for the usage matrix:

Here \(\top\) denotes matrix transpose and the fraction is evaluated for each entry. With positive denominators, it preserves nonnegativity; the corresponding alternating updates do not increase the squared-error objective. A falling objective alone does not certify the globally best solution. This equation is a teaching example, not a claim that multiplicative updates are today’s preferred implementation. [R1]

Assumptions and failure modes

The input representation needs a scientific rationale. Mean-centering can create negative entries, which are incompatible with this basic model. Kotliar and colleagues preserve nonnegativity and discuss why their cNMF preprocessing avoids log transformation: adding components after a logarithm corresponds to a different model in the original units. Their choices belong to their workflow. One should not infer that a familiar preprocessing recipe is automatically appropriate for every measurement platform. [R3]

Identifiability is a separate question from reconstruction error. Scaling and relabeling factors are unavoidable ambiguities; more substantial alternative decompositions can also exist. Donoho and Stodden establish sufficient recovery conditions in a particular parts-based generative setting, including separated support and systematic sampling of combinations. Arbitrary omics data are not guaranteed to satisfy those assumptions. In particular, an attractive list of high-weight genes cannot by itself establish that a fitted component corresponds to a unique biological mechanism. [R4]

The chosen number of patterns and repeated-run behavior deserve explicit inspection. Brunet and colleagues use consensus of sample assignments across initializations as a stability aid. Component-level consensus in cNMF instead combines repeated estimates of the patterns themselves. Neither operation supplies a universal, assay-independent number of biological programs. When planning an analysis, specify which conclusions should survive reasonable changes in rank, feature selection and random initialization. [R2] [R3]

Verified applications

Brunet and colleagues applied NMF to cancer microarray data, using metagenes to describe expression patterns and support sample classification. Their examples include leukemia, medulloblastoma and central nervous system tumors; the leukemia analysis examines known subtype structure. The verified contribution relevant here is an early expression-data workflow linking factorization and stability analysis. Its historical comparisons do not establish superiority over contemporary methods or validate an unseen cohort. [R2]

Kotliar and colleagues developed consensus NMF and evaluated identity and activity program recovery in simulations and published brain organoid and visual cortex single-cell datasets. Their procedure filters and groups components from repeated fits before forming consensus programs and estimating usage. Reported applications include cell-cycle and neuronal-activity patterns. These are documented examples of program discovery; a label attached to an inferred program remains an interpretation requiring the biological evidence presented for that setting. [R3]

Possible transfers — speculative

A possible transfer would be to a nonnegative protein-abundance matrix measured across independently collected samples. This is a speculative proposal, not an application validated by the cited papers. Start with a small panel whose units and background correction are understood. Before fitting, write down whether additive contributions are plausible and whether a common technical intensity shift could masquerade as a program. Inspect candidate patterns alongside run order, sample preparation and other relevant experimental variables.

For that proposed pilot, fit a modest range of pattern counts, retain the seed and preprocessing record, and compare matched components across runs. Reserve independent samples for checking whether the discovered patterns remain useful. A sensible success criterion would be stable, interpretable reconstruction that improves on a simple baseline under the same evaluation design. If interpretation changes with a minor preprocessing choice, keep that instability visible rather than selecting the most appealing gene or protein list.

What to try and what to read next

The small exercise is worth completing before running a large analysis. Reconstruct all four columns, perform the tenfold rescaling, and explain why reconstruction stays fixed while raw usage coefficients change. Then remove the observations that contain only A or only B and ask what evidence remains for separating the two patterns. You do not need a new biological dataset to expose these questions. Pencil-and-paper calculations already distinguish a representation, a numerical fit and an interpretation. The toy also contains exact zeros: three entries of the update denominator vanish, so applying the displayed ratio without numerical safeguards would be undefined.

For a first reading session, use Lee and Seung for the objective and update rule, Brunet for an expression-data example, Kotliar for consensus gene programs, and Donoho and Stodden for the conditions behind recovery. Before trusting a program, be able to explain its units, the alternatives that fit almost as well, and the independent evidence supporting its biological name. Those answers make an additive model scientifically useful.

Reading list

• [R1] Primary reading: Algorithms for Non-negative Matrix Factorization. Daniel D. Lee and H. Sebastian Seung (2000). Read sections 2–4 first, then the auxiliary-function argument in section 6. The official proceedings identify this as NIPS 2000, volume 13; no month or day is asserted.

• [R2] Related reading: Metagenes and molecular pattern discovery using matrix factorization. Jean-Philippe Brunet, Pablo Tamayo, Todd R. Golub and Jill P. Mesirov (2004). Read Methods and the leukemia example for the connection between an additive factorization, sample clustering and repeated-run stability. The online publication date is March 11, 2004; the journal issue is March 23.

• [R3] Related reading: Identifying gene expression programs of cell-type identity and cellular activity with single-cell RNA-Seq. Dylan Kotliar, Adrian Veres, M. Aurel Nagy, Shervin Tabrizi, Eran Hodis, Douglas A. Melton and Pardis C. Sabeti (2019). Read Figure 1 and the consensus-factorization description, then Data preprocessing. This is consensus NMF, distinct from the sample-assignment consensus in Brunet et al. The accepted manuscript was published July 8, 2019 and the version of record July 18; the 2018 bioRxiv preprint is the same study.

• [R4] Related reading: When Does Non-Negative Matrix Factorization Give a Correct Decomposition into Parts?. David Donoho and Victoria Stodden (2003). Read sections 2–3 for the identifiability question; section 5 states sufficient conditions in a synthetic image model. The publisher identifies NIPS 2003, volume 16; this is not a biological validation paper.