Detection without prediction: an empirical capability boundary for covariance-driven single-perturbation models

Abstract

Predicting a cell’s response to genetic perturbation from its unperturbed state is a central goal of computational biology. The CIPHER framework derives that response from the gene-gene covariance of control cells (ΔX = Σu) and reports high accuracy. We reproduced CIPHER with its authors’ released code on sixteen single-cell perturbation datasets (2.0 million cells profiled) and identify an empirical capability boundary for covariance-driven single-perturbation models. One axis dominates the control covariance: its leading eigenvector explains 42–87% of the variance, its loadings match the gene-mean vector (|r| = 0.987–0.998), and equal-depth resampling halves its share (0.68 to 0.26). Because the covariance columns are nearly collinear (median cosine 0.987), a random gene’s column matched on mean, variance and detection predicts as well as the perturbed gene’s (scale-only R² 0.437 vs 0.423, p = 0.49), and a single global mode is the strongest predictor (CIPHER uncentered R² 0.266 vs 0.215 for the full model; scale-only full minus global = −0.031 [−0.049, −0.015]). Predictions are therefore nearly identical across perturbations (mean pairwise |cos| 0.93 versus 0.46 for the true responses) and retain no perturbation-specific similarity structure (RSA = −0.027, p = 0.32). The inverse task does recover the perturbed gene (AUC 0.620.99), but only through that gene’s own coordinate: removing it collapses the AUC from 0.93 to 0.37, while removing the covariance changes nothing (p = 0.74). Beneath this axis lies a real, reproducible target-specific structure (split-half reliability 0.49 [0.31, 0.67], n = 12, p = 5 × 10⁻⁴) that transfers across laboratories, cell lines, libraries and stimulation time (0.28–0.49), is organised by PPI, pathways and transcription-factor regulons, and yet is invisible to the covariance/precision operator in every data space and estimator we tested (off-axis predictability −0.004 [−0.012, 0.003], p = 0.52). On the scale-only R² we detect no gene-specific advantage larger than about 0.01–0.05; this bounds the effect we can detect without proving that no gene-specific signal exists. We give a specificity-audit protocol and recommend matched-random-column, equal-depth and globalmode baselines.