Simulating Shared Signal Across Omics Blocks

What Makes a Signal Shared?

To test whether integration helps, a simulation must specify what information is available in each omics block and how that information relates to the outcome. Here the target is a coordinated signal: multiple blocks measure the same underlying subject-level quantity, and that quantity is associated with a continuous outcome. The measurements need not be identical. Each can contain its own noise, and different features can have different signs, while still carrying information about the same latent quantity.

A useful generator makes these relationships explicit. Let \(i=1,\ldots,n\) index subjects, \(m=1,\ldots,M\) index omics blocks, and \(j=1,\ldots,p_m\) index features in block \(m\). The latent signal for subject \(i\) is \(Z_i\). For clarity, suppose subjects are independently sampled, \(Z_i\sim N(0,1)\), and each block measures this same \(Z_i\).

For a clean benchmark, use a common feature loading strength \(\rho\) across blocks and let \(a_{jm}\) be either \(+1\) or \(-1\) for selected signal features. Define a selected set \(S_m\) in each block, with size \(k_m\). Generate features and outcome as follows:

The residual variables \(U_{ij}^{(m)}\) are independent \(N(0,1)\) draws across subjects, features, and blocks. The outcome errors \(\varepsilon_i\) are independent \(N(0,\sigma^2)\), independent of \(Z\) and all feature residuals. Take \(0<\rho<1\), \(\sigma^2>0\), and \(0\leq k_m\leq p_m\). These assumptions give unit marginal variance to every feature: for \(j\in S_m\), the variance is \(\rho+(1-\rho)=1\), and for \(j\notin S_m\) it is 1. The Gaussian assumptions are a convenient simulation choice, not a requirement for the general idea.

Support terminology needs care. \(S_m\) is the selected set in the generator. When \(k_m>0\), \(0<\rho<1\), and \(a_{jm}\in\{-1,+1\}\), every selected feature has a nonzero coefficient on \(Z_i\), so \(S_m\) is also the feature support for the latent signal in that block. A feature outside \(S_m\) has zero coefficient on \(Z_i\). For an outcome-associated signal, also require \(\theta\ne 0\). If \(\rho=0\), selected features do not actually measure \(Z_i\); if \(\theta=0\), \(Z_i\) is not associated with \(Y_i\). In either case, the corresponding part of the intended shared signal is absent. State selected sets and nonzero supports separately whenever a parameter choice allows them to differ.

How the Generator Creates Coordinated Association

For a selected feature, the covariance with the outcome follows directly from the model. The independent residuals have zero covariance with \(Z_i\) and with each other, so

An unselected feature has covariance zero with \(Y_i\) under the stated assumptions. Since both variables have mean zero, the covariance is also their unstandardized association. The sign \(a_{jm}\) changes the direction of a feature's association, not the underlying latent signal it measures. Because \(\operatorname{Var}(Y_i)=\theta^2+\sigma^2\), the population squared correlation between a selected feature and the outcome is \(\rho\theta^2/(\theta^2+\sigma^2)\). This gives two distinct levers: \(\rho\) controls how clearly features measure \(Z_i\), while \(\theta^2/\sigma^2\) controls the outcome signal relative to error.

Why Feature Support Can Affect Recovery

When more selected features measure \(Z_i\), a method may have more measurements from which to recover it. A simple derivation makes this mechanism visible without assuming that an integration method will necessarily benefit. Suppose there are \(L\) selected features in total, all with the same \(0<\rho<1\). For each feature, multiply its value by its known sign \(a_{jm}\), producing a sign-aligned value \(V_\ell\). Since \(a_{jm}^2=1\), each value has the form \(V_\ell=\sqrt{\rho}Z_i+\sqrt{1-\rho}U'_{i\ell}\), where the \(U'_{i\ell}\) remain independent standard normal variables.

Take the average across the \(L\) sign-aligned features. Conditional on \(Z_i\), its mean is \(\sqrt{\rho}Z_i\), and the variance of its residual part is \((1-\rho)/L\), because the residuals are independent. Across subjects, its total variance is therefore \(\rho+(1-\rho)/L\). Its squared correlation with \(Z_i\), a measure of how much of the latent signal this average recovers, is

The calculation shows why increasing feature support can help in this idealized setting: independent measurement noise averages down as \(L\) increases. It is a justification for varying support in a benchmark, not a guarantee that every fitted integration method will improve. In finite samples, estimation, feature selection, tuning, and model specification can change the observed result.

Designing a Support and Strength Experiment

A simulation should vary one design feature at a time when the aim is to understand its effect. To study feature support, hold \(n\), the block dimensions \(p_m\), \(\rho\), \(\theta\), and \(\sigma^2\) fixed, and compare scenarios with different values of \(k_m\). For example, a small-support condition might select two features in each of three blocks, while a larger-support condition selects eight in each block. Keep the total feature dimensions unchanged; features outside the selected sets remain in the data as noise features.

To examine the strength of the outcome association, hold the selected sets and feature measurement strength fixed, and vary \(\theta\) or the ratio of outcome signal variance to error variance. Under this model, the outcome variance is \(\theta^2+\sigma^2\), and the fraction of outcome variance attributable to the latent component is \(\theta^2/(\theta^2+\sigma^2)\). Specify which quantity is being varied: changing \(\theta\) while holding \(\sigma\) fixed changes both this fraction and the outcome scale.

For each replicate, generate \(Z_i\), all feature residuals, and the outcome errors independently according to the model, then use the same generated dataset when comparing candidate analysis strategies. Across replicates, redraw the random quantities and record the seed and design settings. To make support comparisons interpretable, define how selected indices are chosen and keep that rule consistent; for instance, select the first \(k_m\) indices or randomly select \(k_m\) indices under a documented rule. If the selected sets are redrawn, do so independently of the simulated values and apply the same design rule in each condition.

Diagnostics, Comparisons, and Limits

Check the generator before interpreting benchmark results. Across a sufficiently large simulated sample, each feature should have mean near zero and variance near one. For selected features, the empirical feature–outcome covariance should tend toward \(\sqrt{\rho}a_{jm}\theta\); for unselected features, it should tend toward zero. These are population targets, not exact requirements for every finite replicate. Also check that each block uses the intended number of selected features and that the same subject-level \(Z_i\) was reused across blocks.

A useful diagnostic is to inspect the empirical cross-block associations among selected features. If two selected features have signs \(a\) and \(a'\), their population covariance is \(\rho aa'\) under the independent-residual assumptions. Large finite-sample deviations can occur, but systematic disagreement may reveal an implementation error, such as accidentally drawing independent latent values by block. Check the outcome variance against \(\theta^2+\sigma^2\) as well. A signal that is much weaker than intended may reflect a mistaken scale, a zero coefficient, or excessive outcome error.

The simulation compares approaches to using the same information, rather than proving that integration is beneficial in general. Compare an approach that combines blocks with reasonable approaches that use blocks separately, using identical generated outcomes, subjects, and evaluation splits. The earlier material on integration timing and fair comparison targets is relevant to defining those comparisons; the generator itself cannot make an unfair comparison informative. Differences can depend on the fitted models and evaluation design as well as on the shared signal.

This construction is deliberately controlled. Independent Gaussian residuals, a single common latent variable, equal loading strength, and a linear outcome relationship simplify interpretation. Real data may violate these assumptions. A result in this simulation supports a conclusion about the specified data-generating mechanism, not a universal claim about omics integration. More complicated signal structures should be introduced as separate design changes so their effects are not confounded with support or outcome strength.

Common Mistakes

• Calling a set active when its coefficients are zero. If \(\rho=0\), a selected feature has no \(Z_i\) contribution. Require \(0<\rho<1\) for the selected features in this generator, or explicitly distinguish the selected set from the actual nonzero support.
• Calling a latent factor outcome-associated when \(\theta=0\). A shared feature pattern alone does not create an outcome signal. Require \(\theta\ne0\) when the aim is to simulate features associated with the outcome through \(Z_i\).
• Drawing a new latent value for each block. That removes the shared subject-level driver. Draw \(Z_i\) once per subject and reuse it across blocks.
• Assuming more selected features guarantee better benchmark performance. The derivation describes idealized recovery under independent residuals and known signs. A fitted method can behave differently, so evaluate performance over repeated simulations.
• Changing support and strength simultaneously. If both change, a performance difference cannot be attributed to either design feature alone. Use controlled scenarios and report the parameter settings.