Predicting how a genetic perturbation reshapes a cell’s transcriptome is a central aim in computational biology. Previous reports noted that the simple mean response across training perturbations achieves accuracy comparable to specialized models on standard metrics, yet that mean cannot identify which perturbation occurred. The authors analyze this paradox and propose a new modeling approach that explicitly separates shared and gene-specific response components.
The study examines responses to 2,270 CRISPRi perturbations assayed in each of six cell lines. Across this dataset the authors observed that perturbation responses do not form two discrete classes but instead span a spectrum. Some perturbations elicit responses closely aligned to an overall mean response, while others produce more targeted transcriptional changes that deviate from that mean.
Rather than random or idiosyncratic variation, the authors report that a perturbation’s position along the continuum from mean-aligned to targeted is conserved across cell lines. They quantify this conservation with Kendall’s W = 0.59, indicating moderate-to-strong agreement in relative positions of perturbations across the six cell lines. This conservation suggests a reproducible organization to how perturbations manifest at the transcriptome level.
The study links the continuum position of perturbations to protein-interaction information. Using embeddings derived from the STRING protein-interaction network, the authors show that continuum position is predictable from these embeddings, with reported explained variance R^2 = 0.35. This connection implies that known protein interaction relationships capture part of why some perturbations produce broadly shared responses while others are more specific.
To formalize the observed structure, the authors introduce COMPASS, an interpretable linear model that decomposes each perturbation response into two parts: a shared-response component and a gene-specific residual component. The shared-response is represented as a cell-line–wide response vector that is scaled for each perturbation by a perturbation-specific coefficient. The gene-specific component is the residual after removing the scaled shared-response and is modeled separately to capture targeted effects.
The perturbation-specific scaling coefficients that modulate the shared-response are themselves reported to be strongly conserved across cell lines. This means that for a given perturbation, the degree to which it engages the cell-line–wide shared program tends to be similar across different cellular contexts, reinforcing the conserved organization described earlier.
The residual, gene-specific component shows moderate conservation across cell lines and is reported to recover pathway-level programs. In other words, after accounting for the shared-response, the remaining perturbation-specific patterns capture biologically coherent pathway signatures that are meaningful across contexts.
COMPASS was benchmarked against several recent methods for perturbation prediction, including scGPT, CPA, GEARS, GenePert, and State. The authors report that COMPASS achieves superior response accuracy measured by a de-biased Pearson delta of 0.34 compared with ≤0.32 for the comparator models. COMPASS also improves perturbation discrimination, with a reported cosine PDS gain of 0.23 versus ≤0.08 for the other methods.
Based on these findings the authors recast perturbation prediction across cellular contexts as a problem of component-wise inference. Each response is decomposed into the shared component, estimated from cell-line–wide information and perturbation-specific scaling, and the gene-specific residual, estimated to recover pathway-level effects. The approach assigns each source of evidence to the component it can best inform, improving accuracy and perturbation discrimination.
This work is presented as a preprint and has not been certified by peer review. The article reports the dataset size, conservation metrics (Kendall’s W), predictability from STRING embeddings (R^2), and comparative performance metrics for COMPASS versus listed benchmark models. No competing interests were declared by the authors. Additional methodological details, training procedures, and full evaluation protocols are contained in the preprint; specifics beyond the reported summary metrics are not reproduced here.