Graphical Models as Multi-Omics Dependency Maps

Why Represent Omics Dependencies With a Graph?

A multi-omics analysis can involve many measured features: transcripts, proteins, metabolites, and other molecular quantities. A list of correlations can become difficult to interpret because an association between two features may reflect a relationship between them, or may arise because both are related to another feature. A graph offers a compact way to state which variables are connected in a model and, under explicit assumptions, which pairs are conditionally independent.

The key distinction is between a marginal association, assessed without conditioning on other variables, and a conditional dependency, assessed after accounting for specified variables. They answer different questions. A pair can be marginally associated but conditionally independent, or remain conditionally dependent even after other measured features are accounted for.

Definition: A graphical model has a set of nodes $V$, each representing a random variable, and a set of edges $E$ describing relationships specified by the model. In an undirected graph, an edge joins an unordered pair of nodes. For variables $X_i$ and $X_j$, and a set $X_S$ of other variables, $X_i$ and $X_j$ are conditionally independent given $X_S$, written $X_i \perp\!\!\!\perp X_j \mid X_S$, if their joint conditional distribution factors into their separate conditional distributions: $$ P(X_i,X_j\mid X_S)=P(X_i\mid X_S)P(X_j\mid X_S) $$ almost surely.

A node represents a variable, not an individual sample. For example, a node might be the measured expression of one transcript across the samples in a study. An edge represents a modeled relationship between variables, not a link between two samples. The graph’s vertices and edges therefore describe a population-level dependency structure, while the measurements provide evidence used to estimate or assess that structure.

What an Edge Means Depends on the Graph Model

For a precise introductory interpretation, consider an undirected graph with the following pairwise Markov assumption: whenever two distinct nodes have no edge between them, their variables are conditionally independent given all the other variables in the graph. If $X_S$ denotes all variables except $X_i$ and $X_j$, this assumption says:

(i,j)\notin E \quad\Longrightarrow\quad X_i\perp\!\!\!\perp X_j\mid X_S.

Under this convention, an absent edge has a specific interpretation: the model asserts conditional independence given the remaining modeled variables. An edge indicates that the graph does not assert this particular conditional independence. To interpret an edge as evidence of actual conditional dependence, rather than merely as a retained connection in a model, a further minimality or faithfulness condition is needed: the graph should not include an edge for a pair that is in fact conditionally independent given the other nodes.

These statements are about a specified set of variables. If an important common driver is unmeasured or excluded, conditioning on “all the other variables” in the graph does not account for that driver. A graph can thus be a correct description of its chosen variables and still fail to capture dependencies relevant to the underlying biology.

Not every graphical model uses undirected edges. A directed acyclic graph uses oriented edges and a different set of rules for interpreting its structure. The simple undirected rule above should not be applied to a directed graph just by ignoring arrowheads. For an undirected dependency map, the conditional-independence convention must be stated before absent edges are read as conditional independencies.

Worked Example: A Shared Driver Creates Marginal RNA–Protein Association

Suppose $C$ is a cell-composition score, $T$ is transcript abundance, and $P$ is protein abundance, all standardized to have mean zero. Consider a hypothetical model: $$ C\sim N(0,1),\qquad T=C+\varepsilon_T,\qquad P=C+\varepsilon_P, $$ where $C,\varepsilon_T,\varepsilon_P$ are mutually independent and $\varepsilon_T,\varepsilon_P\sim N(0,1)$. This model assumes the same $C$ is measured for each aligned sample, independent errors for the two omics measurements, and normally distributed variables. It is an illustrative population model, not a real study.

Marginally, both $T$ and $P$ contain the shared term $C$. Since the errors are independent of $C$ and of each other, $$ \operatorname{Cov}(T,P) =\operatorname{Cov}(C+\varepsilon_T,C+\varepsilon_P) =\operatorname{Var}(C)=1. $$ Also, $\operatorname{Var}(T)=1+1=2$ and $\operatorname{Var}(P)=2$, so their marginal correlation is $$ \operatorname{Corr}(T,P)=\frac{1}{\sqrt{2\cdot 2}}=0.500. $$ Thus transcript and protein abundance are marginally associated in this model.

Now condition on $C$. Given $C$, the remaining variation in $T$ is $\varepsilon_T$, and the remaining variation in $P$ is $\varepsilon_P$. Those errors are independent. Therefore $T\perp\!\!\!\perp P\mid C$. Under the undirected graph convention, a graph with nodes $C,T,P$, edges $C-T$ and $C-P$, and no $T-P$ edge represents this conditional independence. The marginal correlation of 0.500 is not evidence, by itself, for a direct conditional dependency between the transcript and protein.

Marginal Association Is Not the Same as Conditional Dependency

Marginal association describes a relationship between two variables without adjustment for other variables. Conditional dependency asks whether information about one variable still changes the distribution of the other after conditioning on a specified set. The conditioning set matters: the conclusion $T\perp\!\!\!\perp P\mid C$ does not imply $T\perp\!\!\!\perp P\mid M$ for some different variable $M$, nor does it imply independence after conditioning on nothing.

In the special case of jointly normally distributed variables, zero conditional covariance is sufficient for conditional independence. Outside such distributional assumptions, zero conditional covariance generally establishes only the absence of a linear conditional association, not full conditional independence. This distinction matters when a graph is interpreted as a map of distributions rather than merely a map of linear relationships.

Worked Example: A Transcript–Protein Dependency Remains After Adjustment

Now consider a different hypothetical model with a measured covariate $C$: $$ T\sim N(0,1),\qquad C\sim N(0,1),\qquad P=0.6T+0.8C+\varepsilon, $$ where $T,C,\varepsilon$ are mutually independent and $\varepsilon\sim N(0,1)$. The variables are aligned across samples, and the model assumes the stated linear relationships, independent normal variables, and no additional unmodeled common cause. Here $C$ is included to illustrate adjustment; the model does not establish that changing $T$ would change $P$.

The marginal covariance between transcript and protein is $$ \operatorname{Cov}(T,P) =\operatorname{Cov}(T,0.6T+0.8C+\varepsilon) =0.6\operatorname{Var}(T)=0.6. $$ The variance of $P$ is $0.6^2+0.8^2+1=2$. Thus the marginal correlation is $0.6/\sqrt{1\cdot 2}=0.4243$.

Given $C$, the term $0.8C$ is fixed. The conditional relationship is $P=0.6T+0.8C+\varepsilon$, so the conditional covariance is $0.6\operatorname{Var}(T\mid C)=0.6$. The conditional variance of $P$ is $0.6^2+1=1.36$. Because the variables are jointly normal, this nonzero conditional covariance implies conditional dependence. The graph should therefore retain a $T-P$ edge under the stated undirected convention, even after accounting for $C$. Its presence describes conditional association in this model; it does not establish a causal effect of transcript abundance on protein abundance.

A Short Derivation: Conditioning Can Remove an Association

The chain rule example below shows algebraically how conditioning on an intermediate variable can eliminate a marginal association. Let $T$, $\eta$, and $\varepsilon$ be mutually independent standard normal variables, and define $M=T+\eta$ and $P=M+\varepsilon$. Here $M$ could represent a measured molecular intermediate, though the model is only hypothetical.

The marginal covariance is $\operatorname{Cov}(T,P)=\operatorname{Cov}(T,T+\eta+\varepsilon)=1$. To calculate the conditional covariance given $M$, subtract from each variable its linear prediction based on $M$. For jointly normal variables this gives the conditional covariance; the residual calculation is: $$ T^*=T-\frac{\operatorname{Cov}(T,M)}{\operatorname{Var}(M)}M =T-\frac{1}{2}M, \qquad P^*=P-\frac{\operatorname{Cov}(P,M)}{\operatorname{Var}(M)}M =P-M. $$

Since $\operatorname{Var}(M)=2$, $\operatorname{Cov}(T,M)=1$, and $\operatorname{Cov}(P,M)=2$, expanding the residual covariance gives: $$ \operatorname{Cov}(T^*,P^*) =\operatorname{Cov}(T,P) -\frac{\operatorname{Cov}(T,M)\operatorname{Cov}(M,P)}{\operatorname{Var}(M)} =1-\frac{1\cdot 2}{2}=0. $$ The residuals are jointly normal, so their zero covariance implies conditional independence: $T\perp\!\!\!\perp P\mid M$. Yet $\operatorname{Var}(P)=3$, making the marginal correlation $1/\sqrt{3}=0.5774$. The variables are marginally associated but conditionally independent given $M$. The derivation relies on joint normality for the independence conclusion; without it, the calculation establishes zero conditional covariance, not necessarily conditional independence.

Worked Example: Reading an Absent Edge Carefully

Suppose a hypothetical undirected graph contains four nodes: transcript $T$, protein $P$, cell-composition score $C$, and batch indicator $B$. It has edges $T-C$, $P-C$, $T-B$, and $P-B$, but no $T-P$ edge. Assume the model uses the pairwise Markov rule described above, all four variables are measured on the same samples, and the graph is intended as a minimal representation of their joint distribution.

There are two other modeled variables besides $T$ and $P$, so the missing edge means $$ T\perp\!\!\!\perp P\mid (C,B). $$ It does not mean $T$ and $P$ are marginally independent. For instance, $C$ could be a shared driver as in the first example, making a marginal association compatible with the missing edge. Nor does the graph say that the batch effect is biological: the $B$ edges represent modeled dependencies, which may reflect technical as well as biological structure.

The conclusion also depends on the assumptions. If the graph is not minimal, a retained edge might not correspond to actual conditional dependence. If relevant variables are omitted, the stated conditioning set does not account for them. And if the graph was estimated from a small or noisy data set, its missing edge may reflect limited evidence rather than a well-supported independence. The graph’s structure is therefore an interpretable model claim, not a guarantee about the data-generating system.

Assumptions, Diagnostics, and Alternatives

Before interpreting an omics dependency graph, check that its basic setup is defensible. Each node should correspond to a clearly defined random variable, and measurements intended to be compared should be aligned to the same observational units. The variables included in the graph should be stated; a conditional independence claim is always relative to that set. The graph type and its edge semantics must also be clear. For the undirected interpretation used here, absent edges imply the specified conditional independencies only under the pairwise Markov assumption.

Diagnostics should target both the data and the interpretation. Investigate whether sample alignment, missingness, technical batches, or cell composition could create or obscure dependencies. Ask whether the assumed distributional relationship is appropriate: a linear association check cannot establish full conditional independence for arbitrary distributions. Check whether conclusions change under reasonable preprocessing or variable inclusion choices. A very dense graph can be hard to interpret, while a sparse graph can omit relationships; neither appearance alone demonstrates that the model is correct.

A correlation matrix is a useful alternative when the goal is to summarize pairwise marginal linear associations. It does not, by itself, encode conditional independence. A graph makes stronger, conditioning-dependent claims, but those claims require a graph model and assumptions. Directed graphs can express a different structure and use different interpretation rules; they should not be treated as undirected dependency maps. Detailed estimation methods and graph-specific separation rules require their own treatment.

Finally, an edge is not a causal conclusion. Even when a conditional dependency persists after adjustment, unmeasured variables, measurement error, selection, and model misspecification can affect the relationship. A graph can organize hypotheses about dependencies and highlight candidate cross-omics relationships, but causal interpretation requires additional design and assumptions.

Common Mistakes

Reading a marginal correlation as a direct edge. A pair may be associated because both depend on a third variable. Check the relevant conditional relationship before calling it a direct dependency.
Calling an absent edge proof of independence. The implication requires the graph’s Markov assumption, an appropriate variable set, and adequate support for the fitted graph.
Treating every edge as causal. An edge describes a modeled dependency, not a confirmed biological mechanism or intervention effect.
Ignoring graph type. An undirected edge and a directed edge do not have interchangeable interpretations. State which model is being used.
Equating zero correlation with independence. This implication holds for jointly normal variables, but not for arbitrary distributions.

Key takeaway: In an undirected graphical model with the pairwise Markov assumption, a missing edge asserts conditional independence given the other modeled variables. A present edge represents a modeled conditional relationship, subject to the model’s assumptions. Neither marginal association nor an edge alone establishes causation.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

In the shared-driver model, why can $T$ and $P$ have a marginal correlation of 0.500 but no edge between them?
What does a missing $T-P$ edge assert in an undirected graph with other nodes $C$ and $B$, under the pairwise Markov assumption?
In the direct-dependency example, calculate the conditional variance of $P$ given $C$. Does the nonzero conditional covariance support conditional independence or dependence?
Why is a nonzero edge not sufficient evidence that transcript abundance causes protein abundance?
For arbitrary, non-normal variables, what does zero conditional covariance establish, and what does it not necessarily establish?

Solutions. (1) Both variables contain the shared driver $C$, creating marginal covariance $1$, but their remaining errors are independent after conditioning on $C$; the model therefore has no direct conditional dependency between them. (2) It asserts $T\perp\!\!\!\perp P\mid(C,B)$, assuming the graph’s Markov rule is appropriate. (3) The conditional variance is $0.6^2+1=1.36$. The conditional covariance is $0.6$, and joint normality makes this evidence of conditional dependence. (4) The graph represents dependency, not an intervention; confounding, measurement error, and other features of the data-generating process can produce an edge without a causal effect. (5) It establishes zero conditional covariance, but without additional distributional assumptions it does not necessarily establish conditional independence.