Why Represent Molecular Chains as Networks?
A genetic variant may be associated with a transcript, the transcript with a protein, and the protein with a metabolite. A list of pairwise associations does not, by itself, describe this as a chain: it does not specify a proposed direction, distinguish a direct link from an indirect one, or show which variables were adjusted for when an effect was estimated. A directed network makes those choices explicit.
The network is a statistical representation of a scientific hypothesis, not a picture that proves the hypothesis. Its structure determines which variables enter each edge model. Fitted coefficients then quantify conditional associations for the proposed edges. Interpreting those coefficients as causal effects requires additional assumptions, which must be stated rather than inferred from the appearance of the graph.
Nodes, Edges, and Structural Equations
For a molecular chain, a possible ordering is genotype \(G\), transcript abundance \(R\), protein abundance \(P\), and metabolite abundance \(M\). The chain \(G\to R\to P\to M\) is acyclic: following its arrows never returns to a node already visited. The ordering matters because each variable is modeled using its proposed parents, not merely by regressing every variable on every other variable.
A linear structural equation model assigns one equation to each non-root node. A root node has no parents in the proposed graph. For node \(Y_j\), one common model is
where \(\operatorname{Pa}(j)\) is the set of parents of \(Y_j\), \(\beta_{kj}\) is the coefficient for the edge \(Y_k\to Y_j\), and \(\varepsilon_j\) is the node-specific residual. In applications, measured covariates may also enter the equation. An edge coefficient is conditional on the other included parents and covariates; it is not generally the same as the unadjusted association between the two variables.
A typical model uses independent individuals, a correctly specified parent set, a linear additive conditional mean, and residuals with mean zero conditional on the included predictors. For a causal interpretation of an edge, stronger assumptions are needed: relevant common causes of that parent and child must be controlled, the proposed direction must be correct, and the variables and model must represent the scientific process adequately. Statistical association alone does not establish these conditions.
Estimating and Interpreting an Edge
For each node with parents, fit a regression of that node on its proposed parents and any prespecified covariates. Under the stated linear model, the fitted coefficient for a parent estimates its conditional edge effect. For example, in \(P=\beta_0+\beta_{GP}G+\beta_{RP}R+\varepsilon_P\), \(\beta_{RP}\) describes the expected difference in protein abundance per one-unit difference in transcript abundance, comparing observations with the same modeled genotype dosage.
The units follow from the variables. If \(R\) is measured in log2 transcript abundance and \(P\) in log2 protein abundance, then \(\beta_{RP}\) is log2 protein-abundance units per log2 transcript-abundance unit. A coefficient does not become a causal effect just because it is attached to an arrow; that interpretation additionally depends on the causal assumptions.
Worked Example: Translating a Chain Into Edge Models
Suppose a hypothetical study measures genotype dosage \(G\) in allele copies, transcript abundance \(R\) and protein abundance \(P\) in log2 units, and metabolite abundance \(M\) in log2 units. The proposed graph is \(G\to R\), \(G\to P\), \(R\to P\), \(G\to M\), \(R\to M\), and \(P\to M\). Write the node equations and identify the parents of each molecular node.
The parent sets are \(\operatorname{Pa}(R)=\{G\}\), \(\operatorname{Pa}(P)=\{G,R\}\), and \(\operatorname{Pa}(M)=\{G,R,P\}\). A corresponding model is
Each edge coefficient is estimated in the equation for its child. Thus \(\beta_{RP}\) is estimated in the regression of \(P\) on both \(G\) and \(R\), not by a regression of \(P\) on \(R\) alone. Its units are log2 protein-abundance units per log2 transcript-abundance unit, conditional on genotype dosage. This specification assumes the proposed graph is acyclic and that these are the chosen parent sets; the equations alone do not establish that the arrows are biologically correct.
What Conditional Regression Establishes
The reason to condition on the other parents follows directly from the structural equation. Suppose the equation for \(Y\) is \(Y=\beta_0+\beta_X X+\beta_Z Z+\varepsilon\), and \(E(\varepsilon\mid X,Z)=0\). Then
Holding \(Z\) fixed and comparing \(X=x+1\) with \(X=x\) gives
This justifies interpreting \(\beta_X\) as a conditional mean contrast in the specified model. If \(Z\) is an important common cause and is omitted, the unadjusted regression generally need not estimate that contrast. If an important cause is unmeasured, adding other measured parents may not remove the resulting bias. The regression identifies the conditional association represented by its model; causal interpretation depends on the assumptions connecting that model to the underlying process.
Worked Example: Estimating Edge Effects and a Chain Product
In a hypothetical sample, fitted regressions give the following estimates. Genotype dosage \(G\) is measured in allele copies; \(R\), \(P\), and \(M\) are in log2 units. The model for \(P\) includes both \(G\) and \(R\), and the model for \(M\) includes \(G\), \(R\), and \(P\).
| Edge | Conditional coefficient estimate | Standard error |
|---|---|---|
| \(G\to R\) | 0.40 log2 \(R\) units per allele copy | 0.08 |
| \(R\to P\), conditional on \(G\) | 0.50 log2 \(P\) units per log2 \(R\) unit | 0.10 |
| \(P\to M\), conditional on \(G,R\) | 0.30 log2 \(M\) units per log2 \(P\) unit | 0.12 |
The estimated effect along the specific chain \(G\to R\to P\to M\) is the product of its estimated edge coefficients:
The units cancel across the intermediate molecular measurements, leaving an estimated 0.0600 log2 \(M\) units per allele copy for this chain product. This is a point estimate under the specified linear network. It is not, on its own, a confidence interval or a test of the whole pathway, and interpreting it causally requires the model and identification assumptions to hold. In particular, the product is meaningful as a proposed chain effect only if the edges and their directions adequately represent the process.
Each coefficient also has its own sampling uncertainty. For example, an approximate 95% interval for the \(P\to M\) coefficient is \(0.30\pm1.96(0.12)\), or approximately \(0.065\) to \(0.535\) log2 \(M\) units per log2 \(P\) unit. This interval excludes zero under the fitted model, but does not independently verify that protein abundance causally changes metabolite abundance. A pathway-level uncertainty calculation must account for uncertainty in all the relevant edge estimates.
Correlation Does Not Determine the Network
A pair of molecular measurements can be correlated even when the proposed network contains no direct edge between them. A shared upstream variable can induce that association. Therefore, adding an arrow whenever two measurements correlate risks turning common-cause associations into unsupported direct links. Conversely, a weak marginal association need not rule out a conditional edge, because other pathways can mask or alter the marginal pattern.
Worked Example: A Correlation With No Direct Molecular Edge
Consider a hypothetical population model with standardized genotype \(G\), transcript \(R\), and metabolite \(M\):
Assume \(E(G)=0\), \(\operatorname{Var}(G)=1\), both residuals have mean zero and variance \(0.75\), and the residuals are mutually independent and independent of \(G\). The graph has \(G\to R\) and \(G\to M\), but no edge \(R\to M\).
The variances are \(\operatorname{Var}(R)=0.50^2+0.75=1\) and \(\operatorname{Var}(M)=0.40^2+0.75=0.91\). The covariance is
Thus the correlation is \(0.20/\sqrt{(1)(0.91)}\approx0.210\), despite there being no direct \(R\to M\) link in the model. In a regression of \(M\) on \(R\) alone, the population slope is \(\operatorname{Cov}(R,M)/\operatorname{Var}(R)=0.20\). But in the regression of \(M\) on \(G\) and \(R\), the coefficient of \(R\) is zero under this model: once \(G\) is held fixed, the independent residual variation in \(R\) carries no information about \(M\). The marginal correlation is explained by their shared upstream cause, not by a direct molecular edge.
This calculation distinguishes a correlation from a conditional edge estimate under the specified model. It does not show that genotype is always the only common cause in real data, or prove that the assumed graph is correct.
When Is an Edge Supported?
An edge is statistically supported when its estimated conditional coefficient is sufficiently incompatible with the null value of zero under the chosen inferential procedure. For a single coefficient, a confidence interval excluding zero is one common indication. The conclusion is specific to the model, sample, and adjustment set. If many edges are tested, the analysis should account for the multiplicity of those tests; otherwise, some apparently supported edges may arise by chance.
Statistical support for a coefficient does not settle whether the edge is causal or whether the network is uniquely determined. Different directed networks can sometimes imply the same observed associations, and an omitted common cause can make a non-edge pair look conditionally associated. Evidence about direction may require external biological knowledge, interventions, temporal information, or additional design assumptions. In reporting results, distinguish clearly among a hypothesized edge, an estimated conditional association, and a causal claim.
Common Mistakes
- Drawing an edge from a correlation alone. A correlation between transcript and metabolite does not show a direct transcript-to-metabolite link. Check whether a proposed shared parent could account for the association, and estimate the edge conditional on the other modeled parents.
- Calling a conditional coefficient causal without qualification. Regression adjusts only for variables included in the model. Causal interpretation also requires appropriate direction, control of relevant common causes, and a suitable model.
- Using the wrong parent set. Estimating \(P\) on \(R\) alone does not estimate the \(R\to P\) coefficient in a graph where \(G\) also points to \(P\). Fit the child on its proposed parents.
- Treating a supported edge as proof of an entire pathway. A coefficient test addresses one modeled edge. A multi-edge chain involves all its component estimates and assumptions, as well as uncertainty in the combined effect.
- Ignoring units. State the change in the child per unit change in the parent, conditional on the other modeled predictors. This makes clear what the coefficient measures.
Check Your Understanding
Answer each question by referring to the specified network and its assumptions.
- In \(G\to R\), \(G\to P\), and \(R\to P\), which variables belong in the regression used to estimate the \(R\to P\) coefficient?
- Why can \(R\) and \(M\) be correlated even if the proposed graph has no direct edge between them?
- A fitted edge coefficient is 0.25 log2 protein units per log2 transcript unit, conditional on genotype. State what this coefficient describes and one assumption needed for a causal interpretation.
- What does a confidence interval for one edge coefficient that excludes zero support, and what does it not establish?
- Why is a chain product not enough by itself to conclude that a molecular pathway is causal?