Soft Thresholding And Weighted Adjacency

From Correlations to Weighted Edges

A correlation network represents genes as nodes and their pairwise expression relationships as edges. Hard thresholding keeps an edge only when a correlation measure passes a cutoff. Soft thresholding instead assigns a numerical weight to each pair, so stronger similarities receive larger weights while weaker similarities are reduced rather than automatically removed. This tutorial develops the power transformation used to make that weighting explicit.

The power parameter controls how sharply the weights distinguish stronger from weaker correlations. It does not change the estimated correlations themselves, and it is not a measure of statistical significance. Its role is to transform a chosen nonnegative measure of pairwise similarity into an adjacency weight.

The construction assumes that the input similarities are valid correlation estimates, that \(\beta\) is positive, and that the same transformation is applied to every gene pair. The correlation estimates may be sample estimates, or may come from an estimation procedure such as shrinkage; the power transformation does not itself correct bias, confounding, or sampling variability in those estimates. The resulting network is therefore conditional on both the correlation estimates and the selected power.

For \(0<s_{ij}<1\), increasing \(\beta\) decreases that pair's weight. A similarity of exactly 1 remains 1 for every positive power, and a similarity of 0 remains 0. Thus the power does not simply scale all edges by the same amount: it changes the relative contrast across different similarity levels.

Why the Power Changes Contrast

The contrast can be seen by comparing two pairs. Let \(s_1\) be the stronger similarity and \(s_2\) the weaker, with \(1\ge s_1>s_2>0\). Their adjacency weights are \(s_1^\beta\) and \(s_2^\beta\). The ratio of the stronger weight to the weaker weight is

Because \(s_1/s_2>1\), this ratio increases as \(\beta\) increases. The stronger pair therefore gains weight relative to the weaker pair. This is a statement about relative weight: both weights can decrease in absolute terms even as their ratio grows.

A short derivative gives another view of the transformation. For a fixed similarity \(s\) strictly between 0 and 1, the weight is \(a(\beta)=s^\beta\). Differentiating with respect to \(\beta\) gives

Since \(\log(s)<0\) when \(0<s<1\), this derivative is negative: the weight decreases as the power rises. For two positive similarities, the derivative of their log weight ratio is

This proves that the ratio favoring the stronger similarity increases with the power. The result does not say that every pair becomes more distinguishable in every absolute sense; it specifically describes the relative weights for a pair of fixed similarities.

Computing a Weighted Adjacency Matrix

For a collection of genes, apply the transformation to each off-diagonal correlation. If the correlation matrix is symmetric, the resulting adjacency matrix is symmetric too, because \(|r_{ij}|=|r_{ji}|\). The diagonal is set to zero by convention for a network without self-edges. The following hypothetical correlations show the calculation for three genes.

Choosing and Comparing Powers

The numerical effect of a power is easiest to see across several correlations. For \(\beta=1\), the weight equals the similarity. Larger powers reduce values strictly between 0 and 1, with stronger similarities reduced less severely than weaker ones.

A power should be reported alongside the correlation convention and the resulting adjacency definition, because the same correlations yield different edge weights under different powers. In practice, the power is a modeling choice, not a quantity estimated by the correlation formula. Its consequences should be inspected across plausible values and in the context of the intended downstream analysis. A larger power can make weak and moderate similarities contribute very little, even though their weights remain mathematically nonzero.

The construction is related to, but distinct from, hard thresholding. A hard-threshold adjacency assigns an edge according to whether a chosen similarity crosses a cutoff; pairs on opposite sides of that cutoff can receive sharply different treatment despite having nearly identical similarities. Power-based soft adjacency retains a graded weight for each positive similarity and changes the degree of contrast smoothly through \(\beta\). Hard thresholding is useful when an explicit cutoff and a sparse set of retained edges are the goal. Soft weighting avoids that particular discontinuity, but it does not guarantee that all edges will be practically influential: a weight near zero may have negligible effect in later calculations.

Diagnostics, Failure Modes, and Interpretation

A sound calculation is simple; a sound interpretation requires checking how its inputs and choices behave. Before using a weighted network, inspect the correlation estimates and transformed weights rather than relying on the power alone.

• Check the input range and symmetry. Correlations should be within \([-1,1]\), and pairwise estimates should be symmetric. Values outside this range or an asymmetric matrix indicate an upstream data or calculation problem.
• Inspect the weight distribution. Summarize or plot the off-diagonal weights. If a chosen power leaves almost every weight near zero, the network may be dominated by a small number of high-correlation pairs. That may be intended, but it should be visible and reported.
• Check sensitivity to the power. Recalculate weights over a reasonable range of \(\beta\). If conclusions based on the network change markedly with small power changes, they depend strongly on this modeling choice.
• Remember estimation uncertainty. A correlation estimated from few samples can be noisy. Raising it to a power does not make it more reliable; it transforms the estimate, including its sampling error.
• Watch for numerical underflow in computation. For large \(\beta\) and small \(s_{ij}\), a computer may represent \(s_{ij}^{\beta}\) as zero because it is below numerical precision. Such a stored zero can be a computational approximation, not an exact zero in the mathematical definition.

A common mistake is to say that a larger power “increases strong correlations.” For any fixed similarity between 0 and 1, increasing the power actually decreases its weight. The correction is that a larger power increases the relative contrast between a stronger and weaker similarity: the stronger weight decreases less, proportionally, than the weaker one. Another mistake is to interpret a high adjacency weight as a significant or causal relationship. It is only the transformed magnitude of an estimated correlation; significance and causal interpretation require different reasoning and evidence.

Finally, this definition uses absolute correlation, so positive and negative correlations with the same magnitude receive equal weights. That is a property of the specified transformation, not a universal rule for co-expression networks. When direction of association is important, the choice of how correlation signs enter adjacency must be made explicitly rather than inferred from these weights.