Do Gene Triples Matter? Condensed Expression Hypergraphs for Perturbation Prediction

Models that predict how a cell's gene expression changes when genes are knocked out or switched on need some prior notion of which genes "work together." Today that prior comes from pairwise networks: curated protein-interaction graphs, co-expression graphs, or ontology similarity. These networks fail exactly where predictions are most needed. Roughly 22% of human genes have too little curated coverage, and on those genes a strong operator model's zero-shot accuracy collapses. A gene's partners can also be read straight from single-cell RNA sequencing (RNA-seq), the measurement of which genes are active in each individual cell, with no curation at all. Treat every cell as a hyperedge, a group linking all the genes active in it at once. This view already exists, but it carries a hidden question that nobody has tested. Pairwise co-expression and principal components depend only on gene–gene correlation. A hypergraph built from cells also records three-way facts: how often genes a, b and c are active together. Those three-way facts are what an epistatic interaction is about, meaning a case where knocking out a and b together does something to target c that neither knockout does alone.

This note proposes to test whether that higher-order co-activity carries information about perturbations beyond what correlation carries. It also proposes to solve the computational problem that has kept the idea from being tested. A cell-as-hyperedge hypergraph over a full transcriptome has tens of millions of "fat" hyperedges, each containing thousands of genes. That size defeats attention-based and sheaf-based hypergraph networks. We condense the cells into a few thousand weighted synthetic hyperedges, chosen so that the condensed set preserves gene-pair and gene-triple co-activity statistics. We then learn gene embeddings from it and plug them into an existing perturbation operator in place of curated-network embeddings.

Core Claim

The claim is falsifiable and has two parts. The first part is that gene embeddings learned from a condensed expression hypergraph will improve combinatorial-perturbation and genetic-interaction prediction, especially for poorly annotated genes. The second part is that this gain disappears when the same hypergraph is built from synthetic cells that match the real gene–gene correlation but have no real higher-order structure. If the gain survives that null, higher-order structure is doing the work. If it does not survive, the honest conclusion is that a cheap correlation embedding is enough. That would also be useful to know, because it would tell the field to stop building expression hypergraphs for this task.

Relation to Prior Work

The gene-as-node, cell-as-hyperedge construction comes from HyperG-VAE, which infers gene regulatory networks with hypergraph self-attention. That attention is quadratic in gene count, which limits it to highly variable gene subsets. HyperG-VAE never asks whether its higher-order view beats correlation, and it does not touch perturbation prediction. We reuse its construction but replace its attention with condensation plus sparse propagation, and we aim it at a different task. HYFA (journal version) is the other expression-hypergraph model. It links individual, tissue and metagene for imputation and sidesteps dimensionality by compressing genes into metagenes, which are not identifiable across runs. We keep all genes as named nodes and compress cells instead, so embeddings stay tied to real genes. High-Order Correlation Integration motivates the question. It shows that correlation-of-correlation features help clustering, but its "higher order" means iterated matrix powers of pairwise correlation, which is still second-moment information. Our covariance-matched null is designed to separate exactly that from true multi-way co-activity.

The condensation method adapts AHGCDD. AHGCDD condenses hypergraphs for node classification using anchor-generated hyperedges and a training-free matching loss based on maximum mean discrepancy, a kernel distance between two distributions. It is about 40× faster than trajectory matching on MAG-PM (27.8 s versus 1,111 s) and runs on Yelp, where HG-Cond runs out of memory. AHGCDD condenses nodes, targets classification, and its edge generation is quadratic in the number of synthetic nodes. We keep the nodes fixed (all genes), condense hyperedges, and match continuous co-activity statistics rather than class prototypes. To our knowledge that is a new use of condensation.

The perturbation side supplies the testbed and the strongest competitors. GEARS predicts multi-gene perturbation outcomes from pairwise co-expression and Gene Ontology graphs. DYNAMO represents knockouts as multiplicative low-rank operators acting on Node2Vec embeddings of STRING and ENCODE networks. DYNAMO is the direct host for our embeddings. It also states the motivating failure: its zero-shot ρΔ drops from 0.283 to 0.108 for genes lacking network coverage. DYNAMO's "regularization paradox" is a constraint on our design. An unconstrained MLP fits better in-distribution but transfers worse, so we must report transfer and not only in-distribution fit. scDEBART challenges the whole premise. It uses no graph prior, pretrains on log-fold changes from 66.6M cells, and reaches an enrichment factor of 11.96, against 1.74–2.99 for GEARS and scGPT. We address it directly below rather than treating graph priors as obviously necessary.

Research Program

The work starts with construction. From control (unperturbed) cells, we binarize each cell's denoised expression to the set of genes active above a gene-specific threshold. Each cell becomes a hyperedge over all roughly 20,000 genes. The incidence matrix is sparse, with nonzeros proportional to total active gene–cell pairs, but it has tens of millions of columns when built from the CELLxGENE corpus.

Condensation comes next. We learn $M'$ synthetic hyperedges (target $10^3$–$10^4$) with nonnegative weights. Each is generated around an anchor gene set, following AHGCDD's anchor idea, with a learned sparsity threshold. The objective matches co-activity on sampled gene sets:

$$\mathcal{L} = \sum_{k=2}^{3} \lambda_k \, \mathbb{E}_{S \sim \mathcal{D}_k} \big( \hat{c}(S) - c(S) \big)^2 + \lambda_{0}\,\mathrm{MMD}(\hat H, H).$$

Here $c(S)$ is the fraction of real cells in which every gene in set $S$ is active, and $\hat c(S)$ is the weighted fraction of synthetic hyperedges containing $S$. The distribution $\mathcal{D}_k$ samples size-$k$ gene sets, enriched for sets involving perturbed genes. The weights $\lambda_k$ balance pair, triple and distribution terms, and the last term is AHGCDD's distribution-matching loss between real and synthetic incidence columns. Uniformly subsampling $M'$ cells already estimates each $c(S)$ with error on the order of $1/\sqrt{M'}$, from standard sampling variance. Learned condensation must beat that rate on triples involving rare genes, or it is not worth having.

Embedding follows. On the condensed hypergraph we train a gene encoder with a self-supervised set-completion task: predict a held-out gene of a hyperedge from the rest. A pairwise objective cannot express this task, because it is inherently multi-way. The encoder uses sparse two-stage node-to-hyperedge-to-node propagation. Its cost is linear in the condensed nonzeros, which avoids the quadratic attention of HyperG-VAE and the quadratic restriction maps that sheaf methods incur on hyperedges this large. We will also report the measured memory and time scaling against the number of source cells.

Finally, the encoder is plugged into the operator model. We substitute these embeddings for DYNAMO's Node2Vec embeddings, keeping its operator architecture, loss and training schedule unchanged, so the embedding is the only variable.

Baselines

The naive reference points are a no-change predictor and an additive predictor that sums single-perturbation effects for double perturbations. DYNAMO reports that multiplicative composition beats addition by 9.6%, so the additive predictor sets the floor for any interaction claim.

The strongest structured competitor is DYNAMO with its original STRING/ENCODE Node2Vec embeddings. It is the right comparison because our method changes only its embedding source. Reported numbers for it are zero-shot ρΔ = 0.283 on unseen K562 genes, 0.108 on genes lacking network coverage, cross-cell-line Pearson 0.576 when transferring from K562 to RPE1, and a reported 2.7× advantage over GEARS on Norman double perturbations (the file does not state the metric behind the 2.7×). We rerun DYNAMO with its released settings on identical splits. Beating it means a higher ρΔ on the uncovered-gene subset, by a margin exceeding seed variance across five seeds, without losing ground on covered genes. GEARS is run under the same Norman splits as a second structured baseline.

The strongest unstructured competitor is scDEBART, with an enrichment factor of 11.96 on Perturb-seq benchmarks and 71.4% top-1 reverse identification. We do not expect a DYNAMO-hosted embedding to beat it outright. The fair comparison is to give scDEBART-style log-fold-change prediction our embeddings as an extra input and test whether enrichment factor rises. Both conditions use the same CELLxGENE cells, so neither side gets more data.

Two ablations isolate the key ingredient. The first is the covariance-matched null: synthetic cells drawn from a Gaussian copula fitted to the real gene–gene correlation and marginal activity rates, then put through the identical pipeline. The second is a pairwise embedding computed from the same real cells, namely principal components of the gene-by-cell matrix and a co-expression graph via clique expansion. A third control replaces learned condensation with uniform cell subsampling at equal $M'$. All variants get the same tuning budget of 30 trials and the same single-GPU budget.

Why not just…?

Why not just use principal components or a co-expression graph from the same cells? This is the most serious threat. A linear degree-normalized hypergraph propagation over cell-hyperedges is close to a rescaled singular value decomposition of the expression matrix. If our encoder learned nothing beyond that, the hypergraph would be redundant. That is why the encoder's objective is set completion and the condensation matches triples, and why the copula null is the central experiment. If PCA embeddings perform as well as ours on uncovered genes and on interaction classification, we will report that correlation suffices.

Why not just use curated gene sets, such as Gene Ontology terms or protein complexes, as hyperedges? They are higher-order, but they share the coverage problem of curated pairwise networks. Poorly characterized genes belong to few or no terms. GEARS already exploits ontology structure and still falls far behind scDEBART in enrichment factor. We include a curated-hyperedge variant and expect it to help covered genes only.

Why not just pretrain a large differential-expression transformer and drop structure entirely? scDEBART suggests this may dominate in-distribution. DYNAMO's results suggest structure matters for transfer: the unstructured MLP retains 30% of its native performance across cell lines versus 51% for DYNAMO. Our K562-to-RPE1 transfer test and the augmentation experiment with scDEBART-style models measure whether hypergraph embeddings add to large pretraining or are subsumed by it.

Why not just subsample cells instead of condensing? At large $M'$ it may suffice. The subsampling control and the triple-error-versus-$M'$ curve show where learned condensation pays off, which we expect for rare-gene triples. If it does not pay off there, we drop condensation in favor of subsampling.

Proposed Experiments

The data are Norman CRISPRa double perturbations in K562, Replogle genome-scale CRISPRi screens in K562 and RPE1, and CELLxGENE control cells for hypergraph construction. Metrics are DYNAMO's ρΔ on held-out genes, split into covered and uncovered genes; mean squared error on the top differentially expressed genes; GEARS-style classification of genetic-interaction types (synergy, suppression, epistasis) on Norman pairs; enrichment factor; and condensation time and peak memory against the number of source cells.

The decisive check is a three-way comparison on uncovered genes and on interaction classification: real-cell hypergraph embeddings, versus copula-null embeddings, versus PCA embeddings. A clear win for real-cell hypergraphs over both, with the gap concentrated on interaction pairs whose triples with target genes deviate most from what correlation predicts, would establish that multi-way co-activity encodes epistasis. A tie would establish the opposite. The secondary check is whether the condensed hypergraph matches full-hypergraph embedding quality at under 1% of the hyperedges, while running where full HyperG-VAE-style attention cannot.