Causal Differential Networks (CDN)

In a Perturb-seq screen the perturbed gene is written on the guide RNA, but in a drug screen or a differentiation protocol it is not. The obvious readout — rank genes by differential expression and take the top hit — works only when the target's own transcript moved most, which is precisely the case where the answer was never in doubt. Knockdowns propagate through cascades that swamp the target, and drugs act on proteins, not mRNA.

Causal Differential Networks (CDN) attacks the problem through causal structure rather than effect size. Given a paired observational and interventional dataset, it infers a noisy causal graph for each, then learns to map the difference between those two graph representations — plus node-level summary statistics — onto the set of variables that were intervened upon. An unobserved change to a mechanism shows up as an observable change to the conditional independencies it induces, so inverting each dataset into its mechanism lets the targets be read off directly. The structure learner and the differential network are trained jointly and supervised, on simulated data and on real Perturb-seq screens.

Menghua Wu, Umesh Padia, Sean H. Murphy, Regina Barzilay and Tommi Jaakkola developed the model in the Department of Computer Science at MIT, posting it to arXiv in October 2024 and publishing it at ICML 2025. It inverts the direction of most single-cell perturbation work: models such as GEARS predict the transcriptional effect of a named perturbation, while CDN starts from the effect and recovers the cause — and does so with no Gene Ontology backbone and no reference interactome, which lets it reach systems where such priors are thin or wrong.

#Key Features

• Targets from graph differences: A differential network consumes the paired graph representations, combined by concatenation or subtraction, and emits a binary intervened/not-intervened label per gene.
• Amortized causal discovery: The structure learner is a trained network rather than an optimizer fit per dataset, so inference runs in seconds on hundreds of variables where classical intervention-target methods take hours on ten.
• No external priors: Predictions come from the data and its inferred structure alone, with no ontology or literature-derived embedding at inference time.
• Cross-dataset transfer: Released checkpoints hold out one cell line at a time, and the Sci-Plex evaluation split is both an unseen cell line and an unseen intervention type.
• Soft interventions: Training simulates "shift" and "scale" interventions alongside hard ones, matching the partial character of real biology.

#Technical Details

Both modules are attention-based. A dataset is featurized as a pairwise correlation matrix plus T = 100 local estimates from the FCI algorithm run over subsets of k = 5 variables; axial self-attention is then applied along outgoing- and incoming-edge dimensions, costing O(N x N^2) rather than the O(N^4) of naive attention over an adjacency matrix, with randomly permuted positional embeddings for equivariance to node labeling. Hidden dimension is 64, with three differential-network layers for the concatenation variant and two for the subtraction variant, under a joint loss combining edge-level graph cross-entropy and target-level binary cross-entropy. The structure learner is initialized from SEA, an amortized causal discovery model from the same group; both modules then train on 8,640 simulated datasets and fine-tune on real screens with only the target loss.

Evaluation covers five Perturb-seq datasets (K562 genome-wide and essential, RPE1, HepG2, Jurkat) and two Sci-Plex chemical perturbation datasets, with candidates restricted to the top 1,000 differentially expressed genes per perturbation. Normalized rank of the true target reaches 0.92–0.98 across the five genetic screens, against 0.45–0.56 for GEARS, GenePT, PDGrapher and linear and MLP baselines, and holding out each cell line costs almost nothing, at 0.91–0.97. On synthetic graphs of 20 nodes it reaches 0.82–0.90 mean average precision on hard interventions, where the strongest causal discovery baseline, DCI, reaches 0.55–0.63.

#Applications

The intended use is narrowing an experimental search space: deconvolving a compound's mechanism of action from its transcriptional signature, choosing transcription factors to test in a cell engineering protocol, or shortlisting causal genes behind a disease-versus-control expression difference. Because the checkpoints transfer to cell lines absent from training, a group with a single observational and interventional pair can score candidates without labeled training data of its own.

#Impact

CDN's broader contribution is methodological: it treats supervised causal discovery as a pretraining objective whose dataset representations serve a downstream biological question, rather than as an end in itself. The framework assumes causal sufficiency — no unobserved confounders — which no transcriptomic assay satisfies, and it models neither cyclic feedback nor time-resolved dynamics. Chemical perturbations remain much harder than genetic ones: across six drugs in two unseen cell lines, the intended protein target fell within the top 100 candidates in three cases. The repository carries no license file, leaving reuse terms for the code unresolved even though the checkpoints and evaluation data are released under CC BY 4.0.