DCAF: Unifying Data Validity, Evaluation Validity and Predictive Modeling for Single-cell Perturbation Simulation

Abstract

Single-cell perturbation profiling is limited by cost and throughput, while computational methods remain constrained by inconsistent evaluation and underperformance relative to simple baselines. Here we present DCAF, a unified framework for single-cell perturbation simulation that that jointly addresses data validity, evaluation validity and predictive modeling. Among its three components, DCAF-data addresses the single-measurement limitation by reconstructing pre- and post-perturbation status using entropy-regularized optimal transport, and is demonstrated on both cell line and multi-cell type datasets. DCAF-criterion evaluates log fold changes in non-silenced genes to improve sensitivity to model error and robustness to zero inflation noise, validated on synthetic datasets. DCAF-model combines JEPA with explicit modeling for regulatory-dynamics modeling for interpretable prediction. Across benchmarks, DCAF outperforms state-of-the-art deep-learning methods, as well as mean and linear baselines in cross-gene, cross-cell-type and cross-region generalization. In adult Mouse brain, DCAF recapitulates known cell-type-specific regulatory heterogeneity of Dgcr8 knockdown at both cell type and gene levels. DCAF thus provides a rigorous and biologically grounded foundation for reliable perturbation simulation across development, disease and therapeutic discovery.