From a fitted factor to a cautious interpretation
A sparse factor can be easier to summarize than a dense one: many feature loadings are exactly zero, leaving a shorter profile to inspect. But sparsity does not make interpretation automatic. A loading is not a causal effect, a sample score is not a clinical outcome, and a list of selected genes is not by itself evidence that a biological pathway is active. Interpretation requires keeping the fitted model, the feature scaling, the factor’s sign convention, and the uncertainty of the selected pattern in view.
Write \(X\) for a centered and, where appropriate, scaled sample-by-feature matrix from one omics block. A common factor representation is \(X \approx SL^\mathsf{T}+E\), where \(S\) contains sample scores, \(L\) contains feature loadings, and \(E\) contains residuals. For factor \(k\), sample \(i\) has score \(s_{ik}\), and feature \(j\) has loading \(\lambda_{jk}\). The fitted contribution of that factor to feature \(j\) for sample \(i\) is \(s_{ik}\lambda_{jk}\). In a multi-block analysis, each block may have its own loading matrix; interpret a loading relative to the block and its scaling.
Interpretation assumes that the chosen factor model and preprocessing give a useful approximation to the measured data, and that feature identifiers, sample alignment, and scaling are correct. The factor must also be meaningfully identified for the intended interpretation: if its loading pattern changes substantially under plausible resampling or modeling choices, a detailed biological label is not well supported. Earlier material on factor stability and uncertainty provides ways to assess that concern; here the emphasis is on interpreting the fitted profile without overstating it.
Read loadings, scores, and signs together
The magnitude of a loading describes how strongly a feature contributes to the factor on the model’s chosen scale. Its sign describes the direction of that contribution relative to the factor orientation. If \(\lambda_{jk}>0\), samples with higher \(s_{ik}\) have a positive fitted contribution from this factor to feature \(j\); if \(\lambda_{jk}<0\), the contribution is negative. This is a statement about the fitted component, not necessarily about the raw feature value alone, since other factors and residual variation also contribute.
The overall sign of a factor is arbitrary: changing \(s_{ik}\) to \(-s_{ik}\) and every \(\lambda_{jk}\) to \(-\lambda_{jk}\) leaves \(s_{ik}\lambda_{jk}\) unchanged. Therefore, “positive” and “negative” only have a stable descriptive meaning after an orientation convention is chosen, such as requiring a specified anchor loading to be positive. Even then, the convention gives a consistent description, not a biological direction like activation or suppression.
Before comparing loading magnitudes, check how features were scaled. If each feature was standardized to unit variance, loadings are on a common standardized-feature scale within that block, although they still depend on the factor normalization and estimation method. If blocks use different scaling or likelihoods, raw loading magnitudes across blocks may not be directly comparable. Scores also depend on factor scaling. In all cases, it is safer to describe the joint loading-score pattern than to treat either number as an independent effect size.
Worked Example: Reading a Sparse Loading Profile
Suppose a fitted factor from standardized expression data has the following nonzero loadings. The model uses \(X \approx SL^\mathsf{T}+E\); all other loadings for this factor are zero.
| Feature | Loading \(\lambda_{j k}\) |
|---|---|
| GENE-A | 0.60 |
| GENE-B | 0.45 |
| GENE-C | -0.50 |
| GENE-D | 0.10 |
For a sample with score \(s_{ik}=2.0\), the fitted factor contributions are \(2.0(0.60)=1.20\) standardized units for GENE-A, \(2.0(0.45)=0.90\) for GENE-B, \(2.0(-0.50)=-1.00\) for GENE-C, and \(2.0(0.10)=0.20\) for GENE-D. GENE-A, GENE-B, and GENE-C have the largest absolute loadings in this list. GENE-D contributes much less on this fitted scale.
The profile supports a description in which GENE-A and GENE-B vary in the same direction as the factor score, while GENE-C varies in the opposite direction, under the current orientation. It does not support saying that high scores cause those genes to change, or that GENE-C is biologically inhibited. If the sign of the entire factor were reversed, all four signs would change while each contribution would stay the same.
Use sample scores to describe who expresses the pattern
Loadings describe which features define a factor; scores describe where samples lie along it. A high score means the sample expresses more of the factor’s signed pattern relative to the fitted score scale. It does not mean that every positively loaded feature is high in raw units, because the model also includes other factors and residuals. A negative score indicates the opposite direction along the same oriented pattern.
A useful interpretation checks whether score differences correspond to a coherent pattern across the selected features, and whether the pattern is supported rather than driven by a few samples. Summaries by a prespecified clinical or experimental group can be descriptive, but a group difference in scores is not evidence of a causal effect. If score uncertainty or stability is poor, emphasize that limitation instead of assigning a confident label.
Worked Example: Connecting Scores to a Feature Pattern
Consider four hypothetical samples and the two larger loadings from the preceding example, GENE-A (\(0.60\)) and GENE-C (\(-0.50\)). The scores are \(-1.5,-0.5,0.5,1.5\), respectively. The factor contributions are the score multiplied by each loading.
| Sample | Score \(s_{ik}\) | GENE-A contribution | GENE-C contribution |
|---|---|---|---|
| 1 | -1.5 | -0.90 | 0.75 |
| 2 | -0.5 | -0.30 | 0.25 |
| 3 | 0.5 | 0.30 | -0.25 |
| 4 | 1.5 | 0.90 | -0.75 |
For sample 4, for example, the calculations are \(1.5(0.60)=0.90\) for GENE-A and \(1.5(-0.50)=-0.75\) for GENE-C. Across these samples, higher scores correspond to a higher fitted contribution for GENE-A and a lower fitted contribution for GENE-C. This describes the factor’s pattern, not the complete observed expression values. With only four samples, the table illustrates the arithmetic; it does not justify a population-level association or a clinical interpretation.
Pathway enrichment: a set-overlap calculation
After defining a feature list, an enrichment analysis asks whether that list overlaps a pathway more than expected under a specified sampling model. For an over-representation test, define a universe of \(N\) eligible features, a pathway containing \(K\) of them, a selected list of \(n\) features from that universe, and an observed overlap \(x\). Under the null model that the list is a uniformly chosen subset of size \(n\), the overlap \(X\) has a hypergeometric distribution. The one-sided enrichment p-value is \(P(X\ge x)\).
The uniform-sampling null is a modeling assumption, not a fact about biological features. The universe should reflect the features that could have been selected and tested, rather than every feature in a broad database by default. The calculation also treats the list membership and annotations as given; it does not account automatically for correlations among genes, uncertainty in loadings, or the fact that the list may have been selected after examining the same data.
Worked Example: Testing a Pathway for Over-Representation
Suppose 1,000 genes passed the analysis’s eligibility filters. A pathway contains 40 of these genes. A list of 20 selected genes from a sparse factor contains 5 genes in that pathway. Thus \(N=1000\), \(K=40\), \(n=20\), and \(x=5\). Under the uniform-subset null, the expected overlap is \(nK/N=20(40)/1000=0.80\) genes.
The one-sided p-value is
The observed overlap of 5 is larger than the null expectation of 0.80, and the calculated p-value is approximately \(0.00079\). Under the stated sampling model, this is evidence that at least this much overlap would be unusual if the 20 genes were a uniformly chosen subset of the 1,000 eligible genes. It is not the probability that the pathway is truly involved, nor does it show that the pathway is activated.
If 30 pathways were tested, multiplicity also matters. For illustration, if this were the smallest raw p-value among 30 tests, the first Benjamini–Hochberg adjusted value would be \(0.00079(30)/1=0.0237\), subject to the usual monotonic adjustment across ranks. The reported evidence should reflect the correction and the full set of tested pathways, not only the most striking raw result.
Separate supported patterns from uncertain ones
A defensible interpretation distinguishes at least three levels of claim. First, a direct description of the fitted result: which features have nonzero loadings, their signs under the chosen orientation, and how scores vary across samples. Second, a statistical summary such as pathway over-representation, with its universe, test, and multiplicity handling stated. Third, a biological explanation, which usually requires evidence beyond the factor fit and enrichment calculation. The third level should be presented as a hypothesis unless independently supported.
Sparse selection can create a particularly tempting mistake: treating zero loadings as proof of no relationship or treating selected features as a definitive biological signature. A zero may result from the model’s penalty or selection rule. A nonzero feature may be selected partly because of sampling variation, correlated features, or modeling choices. Earlier uncertainty and stability analyses are relevant to how much confidence to place in the pattern, but the selected profile itself does not provide certainty.
Interpret enrichment as an annotation-based description of overlap. It is sensitive to pathway definitions, incomplete annotations, gene-set size, the eligible universe, and dependencies among pathway terms. If selected features are strongly correlated, the uniform-subset assumption can be a poor approximation. Enrichment is also distinct from a test of pathway activity: the sign pattern, measurement context, and independent functional evidence all matter. A pathway can be enriched among selected genes without all its members moving together or the pathway being turned on.
Confirm feature preprocessing, score scaling, block membership, and the sign convention used for the factor.
Report selected features, loading signs, and magnitudes in context. Avoid interpreting zeros as proof of no relationship.
Inspect which samples have high or low scores and whether the corresponding fitted contributions are coherent across features.
Specify the eligible feature universe, overlap, null model, and any correction for testing multiple pathways.
Distinguish stable descriptive patterns from tentative biological interpretations, and do not infer causation from the factor fit.
Common pitfalls and alternatives
| Pitfall | Why it fails | Correction |
|---|---|---|
| Calling a positive loading “activation” | Sign is relative to an arbitrary factor orientation and is not a functional assay. | Describe the signed fitted pattern and use functional evidence for activity claims. |
| Calling a zero loading “unrelated” | Sparsity can set a coefficient to zero by design or estimation. | Say that the feature was not selected in this fitted profile. |
| Using all database genes as the enrichment universe | Features outside the analysis’s selection opportunity distort the null comparison. | Use eligible features that could have entered the selected list. |
| Reporting only a raw pathway p-value | Many tested pathways create opportunities for small p-values. | Report the testing scope and an appropriate multiplicity-adjusted result. |
A loading profile can be interpreted directly, compared with another profile after appropriate factor matching, or summarized through annotations such as pathways. Direct inspection preserves feature-level detail but can be hard to summarize. Enrichment is a compact annotation-based summary but depends on its null and database. Neither replaces examining sample scores, scaling, or uncertainty. A dense profile may be preferable when the scientific aim is to represent diffuse contributions rather than select a short list; sparsity is a modeling choice, not a guarantee of clearer biology.
Diagnostics include checking whether a few samples dominate the score range, whether the largest loadings have stable signs and selection across resampling, whether the selected list changes with plausible thresholds or preprocessing, and whether pathway results depend on a small number of annotated features. These checks do not automatically validate the factor, but they can reveal when a polished biological label rests on a fragile pattern. When the interpretation is unstable, report the observed profile and its uncertainty rather than forcing a pathway name.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If every loading and score for a factor is multiplied by \(-1\), what changes and what stays the same?
- A feature has loading \(-0.4\), and a sample has score \(2\). What is that feature’s fitted contribution from this factor?
- Why is a zero loading not sufficient evidence that a feature is biologically unrelated to a factor?
- In an enrichment test, what does the p-value \(P(X\ge x)\) quantify under the uniform-subset null?
- Why should the enrichment universe usually exclude features that could not have been selected?
Solutions. (1) The signs of loadings and scores reverse, but every product \(s_{ik}\lambda_{jk}\), and hence the fitted factor contribution, stays the same. (2) The contribution is \(2(-0.4)=-0.8\) feature-scale units. (3) A sparse estimation procedure can set a loading to zero through its selection rule; that is not proof of no biological relationship. (4) It is the probability, under the specified null sampling model, of observing at least \(x\) pathway members in a list of size \(n\). It is not the probability that the pathway is true. (5) Including ineligible features changes the reference population and therefore the expected overlap; the null universe should represent the features that had an opportunity to enter the list.