Deriving And Interpreting Module Eigengenes

Why Summarize a Module?

A co-expression module can contain many genes measured across the same samples. Looking at every profile at once makes it difficult to describe the module’s overall pattern. A module eigengene compresses those profiles into one score per sample, chosen to capture as much variation among the module genes as a single linear summary can. It is useful as a concise description of shared expression, but it is not a new measurement of a gene, nor does it establish why the genes covary.

The construction depends on how expression profiles are represented. Here, the convention is to center and standardize each gene across samples before calculating principal components. This gives each gene equal variance in the PCA, so genes with large measurement scales do not dominate solely because of their units or spread. Other software or analyses may use different preprocessing; the scaling choice should therefore be reported.

This definition assumes that each included gene has nonzero sample standard deviation; a constant gene cannot be standardized and must be excluded or handled separately. The matrix \(S\) is the sample correlation matrix of the included genes, because the columns of \(Z\) have been standardized. PCA can be calculated algebraically without assuming that samples are independent or normally distributed. Such assumptions matter for particular inferential procedures, not for defining the component. For interpretation, however, samples should be comparable measurements: unaddressed batch effects, cell-mixture differences, or other technical structure can drive the leading component instead of the expression pattern of biological interest.

Deriving the First Principal Component

A linear summary of the standardized gene values in sample \(i\) has the form \(z_i^\mathsf{T}v\), where \(z_i\) is row \(i\) of \(Z\), and \(v\) is a vector of gene weights. Without a constraint on the weights, multiplying \(v\) by an arbitrarily large constant would make the score variance arbitrarily large. PCA restricts \(v\) to have length one and chooses the direction with the largest score variance.

To see why the maximizing direction is an eigenvector, impose the unit-length constraint with a Lagrange multiplier \(\lambda\). At a stationary point of \(v^\mathsf{T}Sv\) subject to \(v^\mathsf{T}v=1\), the gradient of \(v^\mathsf{T}Sv-\lambda(v^\mathsf{T}v-1)\) is zero. Since \(S\) is symmetric, this gives \(2Sv-2\lambda v=0\), or \(Sv=\lambda v\). The variance in this direction is then \(v^\mathsf{T}Sv=\lambda v^\mathsf{T}v=\lambda\). Thus the direction with the largest eigenvalue gives the largest possible variance among unit-length linear summaries.

The first component’s proportion of total standardized gene variance explained is \(\lambda_1/\sum_{j=1}^{q}\lambda_j\). Because each standardized gene has variance one, the total is \(\sum_j\lambda_j=q\). A large proportion indicates that the module’s expression profiles can be summarized efficiently along one direction; it does not by itself establish that the direction is biologically meaningful.

Worked Calculations

Sign, Orientation, and Interpretation

If \(v_1\) is an eigenvector, then \(-v_1\) is also an eigenvector with the same eigenvalue. Replacing \(v_1\) by \(-v_1\) changes every score from \(e_1\) to \(-e_1\), but leaves the component direction, variance explained, and relationships among samples unchanged. Consequently, “positive eigengene score” has no intrinsic meaning until an orientation is selected.

A common convention is to orient the component so that its scores have positive correlation with the average standardized expression of the module genes, when that correlation is nonzero. This makes positive scores tend to mean higher-than-average expression in the module’s overall direction. For a module whose dominant pattern is a contrast between gene groups, however, a simple average can be uninformative or can orient the component poorly. In that case, report the orientation rule and inspect the loadings. Changing the sign for display does not change the PCA result.

Diagnostics, Alternatives, and Limitations

A useful eigengene should summarize a recognizable pattern rather than merely provide a number. Examine the proportion of variance explained, the leading loadings, and the score profile across samples. If the first eigenvalue is close to the second, multiple directions explain nearly as much variation. The first component still exists, but small changes to the data may rotate its direction substantially. When the leading eigenvalue is repeated exactly, the leading direction is not unique: any appropriate unit vector in that tied eigenspace is a valid first component.

Also check whether a few samples dominate the pattern and whether known technical structure aligns with the scores. Plotting gene profiles against the eigengene can reveal genes that do not follow the dominant trend. An eigengene may be driven by batch, cell composition, or an outlying sample; PCA does not automatically distinguish those sources from biological co-expression. If a gene is constant, nearly constant, or has unreliable measurements, standardization can fail or amplify noise, so preprocessing choices require care.

A simple alternative is the average standardized expression, \(q^{-1}\sum_j z_{ij}\). It is transparent and easy to compare across modules, but it gives every gene equal weight and may not capture the direction of greatest variation. The first principal component chooses weights to maximize variance under a unit-length constraint; it can therefore represent a stronger shared pattern, but its weights may be less immediately interpretable and can be sensitive when the leading eigenvalue is not well separated. PCA based on unstandardized centered expression is another alternative: it gives more influence to genes with larger variances. Neither scaling choice is universally correct; the choice should match the scientific meaning of the measurements and be stated explicitly.

The eigengene is a descriptive summary, not a causal explanation or a test of association with a phenotype. Its usefulness depends on the quality of the expression matrix and module definition. A single score necessarily loses information: genes that vary in directions orthogonal to the first component are not represented by that score. When the first component explains little variance, treating it as a complete summary is especially misleading.

Common Mistakes

• Treating the sign as biologically fixed. A negative score is not inherently low expression until an orientation is specified; reversing all loadings reverses all scores without changing the component.
• Calling the eigengene the average expression. It is a weighted principal-component score. It equals a rescaled average only in special cases, such as equal loadings.
• Ignoring standardization. PCA on raw centered expression and PCA on standardized expression answer different questions. State which matrix was analyzed.
• Equating high variance explained with biological validity. A component can capture batch effects or a few influential samples. Inspect its loadings and scores in context.
• Interpreting the score in expression units. With standardized input, the score is a dimensionless weighted sum, not a count or a fold change.