Differential Causal Networks

• Differential causal networks are models that compare causal structures across environments, highlighting changes in edge presence and intervention effects.
• They utilize varied methodologies—from direct difference estimation in linear SEMs to differentiable generative models—to capture both static and dynamic mechanism shifts.
• Applications span biology, neuroscience, and systems governed by PDEs, enabling insights into targeted interventions and regulatory changes.

Searching arXiv for recent and foundational papers relevant to differential causal networks. Search 1: "differential causal networks" Search 2: "causal differential networks intervention targets biology" Differential causal networks denote causal representations of how dependencies, mechanisms, or intervention effects differ across environments, conditions, or times. In one prominent formulation, the problem is to identify which variables were targeted when a biological system moves from an observational distribution to an interventional distribution; in another, the objective is to estimate the edge set of a difference DAG directly from two causal models; in still others, the emphasis is on condition-specific regulatory changes, latent interventions with a shared graph and varying mechanisms, time-varying causal influence, or operator-level changes in PDE-governed systems (Wu et al., 2024, Wang et al., 2018, Faria et al., 2022, Kuskova et al., 21 Mar 2026, Katende, 25 Jun 2025). The literature therefore uses the phrase to cover several closely related ideas: direct estimation of causal differences, causal modeling of paired environments, and structured comparison of networks whose causal semantics are retained under intervention.

1. Conceptual scope and terminological landscape

A differential causal network is not a single canonical object. In the direct-estimation setting of DCI, the target is the difference DAG $\Delta$, defined by edges whose structural coefficients differ between two linear SEMs that share a topological order; this includes both appearance or disappearance of edges and changes in edge weights (Wang et al., 2018). In the biological perturbation setting of Cdn, the goal is instead node-centric: given observational and interventional datasets from the same system, infer the subset of variables whose conditional mechanisms changed, rather than first recovering two exact graphs and then differencing them (Wu et al., 2024).

A second axis of variation concerns whether topology itself is allowed to change. In the latent-intervention framework, all environments share the same DAG while intervention-specific conditional densities vary through target indicators and intervention embeddings; the “differential” content lies in mechanism changes rather than in a changing skeleton (Faria et al., 2022). In CIMLA, differential causal networks are defined through changes in direct causal influence of covariates on an outcome across two conditions, with edge strength derived from a difference of local treatment effects and aggregated by a root-mean-square score (Dibaeinia et al., 2023). In transcriptomic causal network analysis for schizophrenia, “differential” regulation is operationalized primarily as differences in edge presence, mediator loss, and downstream predictive impact between cases and controls (Yazdani et al., 2019).

A third axis concerns dynamical structure. DCNAR treats differential causal networks as changes in causal structure or influence across time, regimes, or groups in multivariate time series, distinguishing static graph discovery from time-varying causal influence estimation (Kuskova et al., 21 Mar 2026). Differential Dynamic Causal Nets extend this idea to EEG-driven neural mass models, where the differential object is a set of node and edge parameter contrasts across groups or time windows (You et al., 29 Jan 2026). By contrast, “differentiable causal discovery” refers to continuous optimization of causal graphs, as in DCD-FG, and should not be conflated with “differential causal networks” in the comparative sense (Lopez et al., 2022). The phrase also appears in a mathematically distinct stream-calculus literature, where it refers to differentiation of causal functions on sequence spaces rather than comparison of causal graphs across environments (Sprunger et al., 2019).

2. Structural formulations and intervention semantics

Most formulations begin from a structural causal model over a DAG. In Cdn, random variables $X=(X_1,\ldots,X_p)$ satisfy structural equations $X_i := f_i(X_{Pa(i)}) + \epsilon_i$, with independent noise, and interventions modify conditional mechanisms on a subset $I \subseteq V$ by replacing $P(X_i \mid X_{Pa(i)})$ with $\tilde{P}(X_i \mid X_{Pa(i)})$ for $i \in I$ (Wu et al., 2024). Hard interventions remove dependence on parents and, in the synthetic setting, use $X_i := z$ with $z \sim \mathrm{Uniform}(-1,1)$; graphically, all incoming edges to $i$ are removed in the mutilated graph. Soft interventions preserve the parent set and functional form but alter parameters, modeled synthetically as $X=(X_1,\ldots,X_p)$0 with $X=(X_1,\ldots,X_p)$1, leaving the DAG unchanged while shifting the induced joint distribution (Wu et al., 2024).

DCI adopts a more restrictive linear-Gaussian setting. Two environments $X=(X_1,\ldots,X_p)$2 and $X=(X_1,\ldots,X_p)$3 are generated by linear SEMs

$X=(X_1,\ldots,X_p)$4

with a shared topological order. The difference network is encoded by $X=(X_1,\ldots,X_p)$5, and an edge belongs to the difference skeleton precisely when the corresponding coefficient changes across environments. Noise variance changes are not represented as edges in $X=(X_1,\ldots,X_p)$6; instead, they enter the orientation stage through residual-variance invariance tests (Wang et al., 2018).

Other formulations preserve graph structure and localize difference to mechanisms. In the latent-intervention mixture model, interventions may be atomic, stochastic, or imperfect/soft, but they “never add parents; they only remove edges,” and the method learns a shared adjacency matrix together with intervention-specific conditional densities under a Dirichlet process mixture (Faria et al., 2022). In time-varying autoregressive models, the differential object is expressed through changes in matrices $X=(X_1,\ldots,X_p)$7 or in impulse responses rather than through a single cross-sectional graph difference. DCNAR writes

$X=(X_1,\ldots,X_p)$8

with a learned structural mask $X=(X_1,\ldots,X_p)$9 constraining admissible influences, so differential analysis may target $X_i := f_i(X_{Pa(i)}) + \epsilon_i$0, $X_i := f_i(X_{Pa(i)}) + \epsilon_i$1, or $X_i := f_i(X_{Pa(i)}) + \epsilon_i$2 across times or groups (Kuskova et al., 21 Mar 2026).

The same comparative logic appears in more specialized domains. In PDE discovery, nodes are spatiotemporal fields and edges are operator-mediated influences; interventions act on operators or coefficients through $X_i := f_i(X_{Pa(i)}) + \epsilon_i$3 or $X_i := f_i(X_{Pa(i)}) + \epsilon_i$4, and differential structure is read off from counterfactual operator removal (Katende, 25 Jun 2025). In EEG-based DDCNs, edges are parameterized transmissions between local neural mass models, and the differential network is the set of directed edges whose parameter contrasts are statistically significant and practically large across conditions (You et al., 29 Jan 2026).

3. Methodological families

A major methodological divide separates difference-first estimation from learn-two-models-and-compare strategies. DCI is explicitly difference-first. It optionally screens for changed nodes and edges through a difference-UG, estimates the difference skeleton by testing invariance of regression coefficients across the two datasets, and orients edges by testing invariance of residual variances. Its two-step logic is designed to avoid separately learning two large DAGs and then differencing them (Wang et al., 2018).

Cdn also avoids a direct joint search over graphs and targets, but in a distinct way. It first computes environment-specific causal graph features with the Sample–Estimate–Aggregate causal featurizer, using global correlation or inverse covariance statistics together with marginal FCI outputs on small subsets. It then concatenates the observational and interventional edge embeddings,

$X_i := f_i(X_{Pa(i)}) + \epsilon_i$5

and applies an axial-attention classifier to predict node-wise intervention targets via

$X_i := f_i(X_{Pa(i)}) + \epsilon_i$6

Training uses binary cross-entropy with AdamW, and the featurizer is amortized and frozen at target-prediction time (Wu et al., 2024).

A third family uses differentiable generative or variational objectives. CGNN instantiates each causal mechanism as a neural network, learns a DAG-structured generator from observational data, and scores candidate graphs through Gaussian-kernel MMD between real and generated samples. Although CGNN is not itself a differential-network estimator, it is a reference point for differentiable causal networks because it learns functional causal models by backpropagation and supports interventions through graph mutilation (Goudet et al., 2017). DCD-FG extends differentiable causal discovery to high-dimensional interventional data by restricting the search space to Boolean low-rank factor DAGs and optimizing a likelihood with continuous acyclicity penalties computed on factor half-squares (Lopez et al., 2022). The latent-intervention method likewise uses differentiable optimization, but here the objective is an ELBO over a Dirichlet process mixture of intervention SCMs with a differentiable acyclicity constraint

$X_i := f_i(X_{Pa(i)}) + \epsilon_i$7

on the shared graph parameters (Faria et al., 2022).

A fourth family replaces explicit graph differencing with condition-specific predictive models plus causal interpretation. CIMLA trains separate Random Forest or neural-network predictors per condition, computes local SHAP attributions, interprets them as local treatment effects under stated assumptions, and defines a differential score

$X_i := f_i(X_{Pa(i)}) + \epsilon_i$8

Edges in the differential causal network are the covariate–outcome relations whose causal influence changes across conditions according to this score (Dibaeinia et al., 2023).

4. Identifiability, guarantees, and computational properties

The theoretical picture is heterogeneous. DCI provides the strongest formal guarantees among the paired-environment graph-difference methods. Under Difference-adjacency-faithfulness, its skeleton estimator is consistent; under Difference-orientation-faithfulness, all edges oriented in the second stage are correctly oriented; and if noise variances are unchanged across environments, the method recovers the full D-DAG (Wang et al., 2018). The appeal of this theory is that assumptions are placed on the differential object rather than on each full DAG separately.

By contrast, Cdn is explicitly a supervised discriminative approach and “does not claim identifiability guarantees.” Its paper offers informal capacity arguments that an axial attention layer can map $X_i := f_i(X_{Pa(i)}) + \epsilon_i$9 to target sets in the hard-intervention case and $I \subseteq V$0 to targets in the soft-intervention case, but correctness depends on the featurizer and on representative supervised data (Wu et al., 2024). The latent-intervention variational method also does not provide formal identifiability theorems, instead relying on acyclicity, mutual faithfulness, intervention diversity, and the shared-graph assumption to justify recovery of the common DAG under sufficient data (Faria et al., 2022).

Differentiable causal discovery methods provide a different kind of theory. CGNN proves existence of an FCM for continuous strictly positive densities on compact support, neural approximation of the corresponding mechanisms, and convergence of MMD to zero under the correct DAG and sufficient capacity (Goudet et al., 2017). DCD-FG proves that the candidate space of acyclic half-squares of factor graphs is exponentially smaller than the full DAG space for fixed factor rank and shows that, under a random growing f-DAG model, the Markov equivalence class size tends to one with high probability as $I \subseteq V$1 grows (Lopez et al., 2022). In PDE operator discovery, the guarantees are strongest of all: exact support recovery is proved under RIP or mutual coherence conditions, and residual bounds imply identifiability of the discovered operator support (Katende, 25 Jun 2025).

Computationally, differential causal network methods trade off structural fidelity against scalability. Cdn reports featurization and inference in minutes on single-cell datasets with up to $I \subseteq V$2 genes and uses axial attention to avoid full $I \subseteq V$3 two-dimensional attention, with row and column attentions each costing $I \subseteq V$4 (Wu et al., 2024). DCI’s cost depends exponentially on the size of screened changed-node sets unless conditioning-set size is capped (Wang et al., 2018). The latent-intervention approach is limited chiefly by the matrix exponential in the acyclicity penalty, with $I \subseteq V$5 complexity (Faria et al., 2022). DCD-FG attains linear-in-$I \subseteq V$6 scaling for fixed factor rank $I \subseteq V$7 because likelihood computation is $I \subseteq V$8 and acyclicity is enforced on factor half-squares (Lopez et al., 2022).

5. Domains of application and empirical findings

In single-cell perturbation biology, Cdn is designed for identifying perturbation targets from paired observational and interventional transcriptomics. On seven single-cell transcriptomics datasets—five Perturb-seq and two Sci-Plex datasets—it “consistently outperforms baselines.” For K562 genome-wide non-trivial perturbations, the reported normalized rank and Recall@20 are $I \subseteq V$9 for Cdn versus $P(X_i \mid X_{Pa(i)})$0 for PDGrapher, $P(X_i \mid X_{Pa(i)})$1 for GEARS, and $P(X_i \mid X_{Pa(i)})$2 for GenePT; on synthetic data it achieves hard-intervention mAP $P(X_i \mid X_{Pa(i)})$3–$P(X_i \mid X_{Pa(i)})$4 and AUC $P(X_i \mid X_{Pa(i)})$5–$P(X_i \mid X_{Pa(i)})$6, while also substantially outperforming alternatives on soft interventions (Wu et al., 2024).

Direct differential graph estimation has been applied to genomics as well. DCI was demonstrated on ovarian cancer gene expression and on naive versus activated T-cell single-cell RNA-seq, where it identified hubs such as BIRC3 and PRKAR2B in apoptosis, THBS2 and COMP in TGF-$P(X_i \mid X_{Pa(i)})$7, and GZMB and UHRF1 in T-cell activation; the paper emphasizes that it additionally provides causal directions relative to undirected differential models (Wang et al., 2018). At much larger scale, DCD-FG was evaluated on a Perturb-CITE-seq melanoma dataset with 218,331 cells, 249 gene interventions, and $P(X_i \mid X_{Pa(i)})$8 genes, where it outperformed NOTEARS, NOTEARS-LR, NOBEARS, and IGSP on interventional NLL and interventional mean absolute error across co-culture, IFN-$P(X_i \mid X_{Pa(i)})$9, and control conditions (Lopez et al., 2022).

Condition-specific gene regulation is another major application. CIMLA uses SHAP-based local treatment effect estimates to infer differential gene regulatory networks and, on SERGIO simulations, is reported to be “more robust to confounding variables and more accurate than leading methods.” Applied to Alzheimer’s disease snRNA-seq from prefrontal cortex, it identified a high-confidence dGRN intersected with PsychENCODE evidence, highlighting regulators such as CREB3, NEUROD6, ELK1, and GATA3 (Dibaeinia et al., 2023). A related observational-but-instrumented strategy was used for schizophrenia, where transcriptomic causal networks in CommonMind dorsolateral prefrontal cortex identified loss of mediator function in genes including GABRA2, LRRTM2, PPM1E, SORT1, and GNAL, and preserved hub influence for TENM3, NRXN3, MYH10, and PEX5L (Yazdani et al., 2019).

Dynamic and biophysical settings extend the concept beyond transcriptomics. DCNAR was evaluated on V-Dem multi-country panels and produced “smooth, sign-consistent, bounded impulse responses and counterfactual trajectories,” with predictive distribution accuracy comparable to Ridge VAR and TV-VAR, and near-nominal $\tilde{P}(X_i \mid X_{Pa(i)})$0 coverage (Kuskova et al., 21 Mar 2026). Differential Dynamic Causal Nets were applied to pediatric epilepsy EEG and found strongest case-control differences in frontal and prefrontal outputs, with significant contrasts in parameters such as $\tilde{P}(X_i \mid X_{Pa(i)})$1, $\tilde{P}(X_i \mid X_{Pa(i)})$2, $\tilde{P}(X_i \mid X_{Pa(i)})$3, and $\tilde{P}(X_i \mid X_{Pa(i)})$4, and strong preictal-to-ictal increases in dynamic causal parameters (You et al., 29 Jan 2026). In scientific machine learning, counterfactual PINNs recovered true operator support exactly on several synthetic PDE benchmarks and achieved, for tumor diffusion, false positive rate $\tilde{P}(X_i \mid X_{Pa(i)})$5, MAE $\tilde{P}(X_i \mid X_{Pa(i)})$6, and Relative $\tilde{P}(X_i \mid X_{Pa(i)})$7 error $\tilde{P}(X_i \mid X_{Pa(i)})$8, while recovering support $\tilde{P}(X_i \mid X_{Pa(i)})$9 (Katende, 25 Jun 2025).

6. Limits, misconceptions, and open directions

Several misconceptions recur in this literature. First, differential causal networks are not the same as classical differential association networks. Classical differential networks compare correlation or partial-correlation structure across conditions and therefore detect association changes, whereas causal differential approaches attempt to localize changes in conditional mechanisms, directed edges, or intervention targets (Wu et al., 2024). Second, differential causal networks are not identical to differentiable causal discovery: DCD-FG is about continuous optimization of causal graphs, and its use for comparative analysis arises by fitting separate models per condition rather than by defining a specialized difference estimator (Lopez et al., 2022). Third, the phrase should not be conflated with the differential calculus of causal functions on streams, which studies upper-triangular causal derivatives and recurrence rules for sequence maps (Sprunger et al., 2019).

The core limitations are equally consistent across papers. DAG assumptions clash with biological feedback, hidden regulators, and context-specific interactions; Cdn explicitly notes feedback loops and latent confounding as limitations, and DCI requires shared topological order and causal sufficiency in its main theory (Wu et al., 2024, Wang et al., 2018). Supervised or variational approaches often lack formal identifiability on real data, with success depending on training distribution, intervention diversity, or representation quality rather than on theorem-level recovery guarantees (Wu et al., 2024, Faria et al., 2022). In predictive-attribution approaches such as CIMLA, the SHAP–causal linkage assumes no unmeasured confounders of covariates and outcome and depends on transferability of observational models to intervention-like queries (Dibaeinia et al., 2023).

Open directions stated in the literature include semi-supervised integration of known targets, domain adaptation across cell types and labs, mixed-graph or partially directed representations that better accommodate confounding and feedback, richer variational families for latent interventions, uncertainty quantification through Bayesian or ensemble methods, and extensions from differential mechanisms to differential topologies (Wu et al., 2024, Faria et al., 2022). In dynamic settings, change-point detection, groupwise comparison of time-varying causal influence, and improved handling of structural uncertainty remain central problems (Kuskova et al., 21 Mar 2026). Across domains, the shared research objective is stable: to preserve causal semantics while comparing systems, so that the discovered differences correspond not merely to altered association patterns but to altered intervention-relevant structure.