Loading…
Causal discovery foundation model that aggregates classical FCI or GIES estimates over small variable subsets into a global gene regulatory graph.
No providers recorded yet. Browse all providers
A Perturb-seq screen can measure thousands of genes and still leave only a few dozen cells per intervention, which is precisely where classical causal discovery breaks down. Constraint- and score-based algorithms are analyzed in the infinite-data limit, must be re-optimized from scratch on every dataset, and are correct only under assumptions nobody can check against real expression data. Running several and comparing outputs is the obvious hedge, and unaffordable when one run takes hours.
SEA — Sample, Estimate, Aggregate — moves the learning off the raw measurements and onto the outputs of those classical algorithms. It draws many small batches of observations, restricts each to a handful of variables, runs a cheap classical algorithm (FCI or GIES) on each subset, and computes one inverse-covariance matrix over all variables. An axial-attention network, pretrained on synthetic structural causal models, reconciles those noisy local estimates and the global statistic into a single N x N causal graph. The premise is that a discovery algorithm's characteristic mistakes are comparable across datasets, so a network can learn to correct them once instead of being refit each time.
Menghua Wu, Yujia Bao, Regina Barzilay and Tommi Jaakkola developed SEA at MIT with Accenture's Center for Advanced AI, releasing it on arXiv in February 2024 and in Transactions on Machine Learning Research in March 2025. Like the catalog's other synthetic-prior models — PerturbPFN, GPFN — its weights never see real biology during training, and it is the base from which CDN initializes its structure learner.
The aggregator is a four-layer, eight-head axial-attention network with hidden dimension 64, trained with cross-entropy over three per-edge classes; predicted graphs are not constrained to be acyclic. Pretraining used 6,480 synthetic datasets — roughly 280 million observations over graphs of 10 to 100 variables, Erdos-Renyi and scale-free topologies with N to 4N expected edges, under linear and neural-network mechanisms with additive and non-additive Gaussian noise, in observational and single-node-intervention regimes. Two aggregators were trained, one over FCI/Fisher-z estimates and one over GIES/BIC, taking about 14 and 16 hours on two RTX A6000 GPUs; the released code carries a third checkpoint fit to SERGIO-simulated mRNA.
On CausalBench K562 Perturb-seq — 622 genes, 162,000 cells, scored against physical STRING edges — SEA reaches F1 0.179 at precision 0.491 in 4 seconds, against 0.142 for NOTEARS in 32,883 seconds and 0.051 for DCDI-G in 16,561; a larger observation batch raises precision to 0.838. On the Sachs flow cytometry benchmark, substituting a kernel independence test and finetuning for correlation lifts mean average precision from 0.23 to 0.49 and gives the comparison's lowest structural Hamming distance, 13. Inference runs over 500-variable graphs on a single 32 GB V100.
The immediate use is reading candidate regulatory structure off a perturbation screen — which genes plausibly act on which — cheaply enough to repeat the analysis under several assumption sets and compare, rather than committing to one algorithm's model of the data. The same machinery applies to protein signaling measured by flow cytometry and to any experiment with paired observational and interventional arms. A pretrained checkpoint transfers with no dataset-specific training, so a group with a single screen and no labels gets a graph in seconds; downstream, CDN reuses that inferred structure to identify which genes a perturbation hit.
SEA's contribution is a recipe rather than a single network: a sampling heuristic, a classical estimator, a summary statistic and a learned aggregator, each replaceable, which lets domain knowledge enter at inference instead of requiring new simulations and a retrain. The published implementations carry real limits: a hard-coded ceiling of 1,000 variables, weak generalization to cyclic data, over-sparse predictions on real datasets, and numerical instability of inverse covariance on large graphs. Causal sufficiency is assumed throughout, which no transcriptomic assay satisfies. And because the weights are fit entirely on simulation, the authors are explicit that trustworthiness in a new domain depends on how faithfully the simulator reflects that domain's graph topologies, functional forms and measurement error.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.