Learning DAGs Under a Prespecified Layer Order

Why Prespecify a Layer Order?

Multi-omics features often have a biologically motivated organization: for example, a regulatory layer may be placed before a transcript layer, which may be placed before a protein layer. Prespecifying that order reduces the number of directed graphs that an estimator must consider. Instead of searching over every possible orientation, it searches only among graphs compatible with the chosen precedence constraints.

This is a restriction on the statistical graph, not evidence that the order is correct. An order can make estimation simpler while excluding plausible edges or imposing an orientation that the observed data cannot verify. The estimated graph must therefore be read in light of the order used to construct it.

The final condition matters. Strictly forward edges cannot form a directed cycle because every step along a path moves to a later layer, and a cycle would have to return to its starting layer. But if edges within a layer are allowed, the layer order alone does not rule out cycles within that layer. Such edges require an additional within-layer acyclicity constraint, such as a prespecified ordering of features or an explicit search over acyclic within-layer graphs.

A Gaussian DAG Model and Local Edge Estimation

A common model for continuous, transformed omics measurements is a linear Gaussian structural representation. Assume \(n\) independent observations of a vector \(X=(X_1,\ldots,X_p)\). For each feature \(X_i\), write

where \(\operatorname{pa}(i)\) is the selected parent set of \(X_i\). The model assumes a directed acyclic graph, linear conditional means, Gaussian errors, and mutually independent errors across features. Variables are taken to be centered here; with uncentered variables, an intercept is included as a nuisance parameter. For a fixed graph, these assumptions imply the factorization \(f(x)=\prod_{i=1}^p f(x_i\mid x_{\operatorname{pa}(i)})\).

Under a strict layer order, the candidate predictors for \(X_i\) are already limited to \(P_i\). If there are no additional constraints linking parent sets, one can estimate each node’s parent set by a regression using only its allowable predecessors. For a candidate parent set \(S_i\subseteq P_i\), fit the regression of \(X_i\) on \(X_j\), \(j\in S_i\), and compare the fit with a complexity penalty. A sparse fitted coefficient \(\hat\beta_{ji}\) corresponds to selecting the directed edge \(X_j\to X_i\).

For example, a BIC-style score for a centered Gaussian regression with residual sum of squares \(RSS_i(S_i)\) is

Lower scores are preferred. This form omits terms common to all candidate parent sets for the same response, including the contribution from estimating its residual variance. The score trades improved fit against the cost of additional edges; it is not a test that an individual edge is true.

The simplification applies to the stated setup. If the model imposes extra joint constraints—for example, a penalty tying coefficients across responses or a restriction on the total number of edges—separate optimizations need not give the optimum under that joint constraint. And if within-layer edges are permitted, independently selecting parents can produce a cycle unless acyclicity is enforced there too.

Why the Likelihood Separates

For a DAG, the joint density factors into one conditional density for each feature. Under the linear Gaussian model, the log-likelihood for observation \(t\) is, apart from constants,

Summing over observations \(t=1,\ldots,n\) groups the terms by response \(i\). The coefficients and variance for \(X_i\) occur only in that response’s regression term. Therefore, when parent sets are chosen independently subject to their allowable sets, the likelihood contribution for one response does not require optimizing the coefficients for another response.

This is the computational reason a known order is useful: it eliminates cyclic candidate graphs when edges point strictly forward, and it restricts each regression to a known list of possible predictors. It does not, by itself, establish that a selected edge represents a direct biological mechanism.

Worked Consequences of the Order

Assumptions, Diagnostics, and Alternatives

A restricted regression search is most defensible when the observations are independent and comparable, the chosen transformation and Gaussian model are adequate, and the layer order is scientifically justified for the intended statistical analysis. The candidate predictors must be measured with enough information to distinguish their conditional contributions; strong collinearity can make selected parents unstable or obscure which of several related features carries the signal. A high-dimensional candidate set also requires appropriate regularization or model selection rather than unpenalized fitting alone.

Check the fitted regressions for patterns that contradict the assumed model: residual distributions markedly inconsistent with Gaussian errors, nonlinear residual trends, heteroscedasticity, or residual dependence across observations. Also compare the selected graph with the allowed-edge list and verify acyclicity explicitly whenever within-layer edges are present. A suspiciously poor fit for a node can indicate that its true predictors were excluded by the order, that relevant features are missing, or that the linear conditional model is inadequate. These checks flag problems; they do not identify a unique correction.

The main alternatives answer somewhat different questions. The earlier material on Graphical Lasso for Joint Omics Networks estimates an undirected conditional-dependence structure rather than an ordered DAG. Nodewise regression and pseudolikelihood are related estimation strategies but do not automatically make directed orientations identifiable. Unconstrained score-based DAG search can consider more structures, but it must handle acyclicity across the search and is generally more computationally demanding. Constraint-based approaches use conditional-independence information and can return a set of graphs when the data do not determine a unique orientation. A prespecified order narrows the search; it does not remove uncertainty about whether that order is appropriate.

A further limitation is that graph selection depends on the score, penalty, sample size, and measurement properties. A small change in the data or penalty can alter selected parents, especially when candidate predictors are correlated. Treat estimated edges as features of a fitted model under stated assumptions, rather than as unqualified biological facts.

Common Mistakes

• Assuming a layer order always guarantees a DAG. It guarantees acyclicity for strictly forward edges. If within-layer edges are allowed, additional constraints are necessary.
• Treating the order as something learned from the same regressions. The procedure described here conditions on a prespecified order. If that order is wrong, the permitted graph set may be wrong too.
• Choosing every predictor that improves fit. Fit almost always improves as predictors are added. A complexity penalty or another justified selection rule is needed to balance fit and graph size.
• Reading a selected arrow as proof of a direct mechanism. The arrow records an orientation permitted and selected under the model and order; it does not, by itself, establish a biological or causal interpretation.