Abstract
<title>Abstract</title> <p>Single-cell perturbation screens connect genetic interventions to cellular states, but combinatorial coverage remains sparse. Existing predictors use labels, gene embeddings, graphs or foundation-model representations to extrapolate responses, yet the response evidence used to support a prediction is often absorbed into latent model representations. Here we introduce Control-Anchored Perturbation Residual Architecture (CAPRA), a response-anchored model for genetic perturbation prediction. CAPRA builds predictions from a sampled control state, an observed or embedding-retrieved single-gene response, learned correction for context and gene-pair composition, and control-scale-calibrated predicted cells. Across diverse perturbation screens, CAPRA achieved the highest mean DEG-masked response-direction correlation among the compared methods in both single- and double-gene tasks. The advantage persisted under perturbed-reference scoring and in affected-gene recovery. By exposing anchor support and learned correction, CAPRA links predictive performance to explicit response evidence for experimentally unmeasured genetic interventions.</p>