A Neural-Network-Free Calibration Matches or Beats Deep Single-Cell Perturbation Response Models Across Four Datasets

Background/Objectives: Deep generative models such as scGen are the standard for predicting how an unseen cell type or species responds to a perturbation under leave-one-group-out (LOCO). We ask whether the deep model's advantage over a one-line linear baseline is learnable without any neural network. Methods: We decompose scGen's edge into two leakage-free analytic pieces (per-type response magnitude and per-gene response direction) and add a per-gene affine moment-matching step. The resulting calibration, AMM-SimWMag, uses no neural network and is CPU-only. We evaluate it on a nine-metric panel across 4 datasets spanning 3 biologies and 2 LOCO axes (cell type and species), using a dataset-stratified test (a Stouffer combination of per-dataset signed-rank tests, Holm-adjusted). Results: AMM-SimWMag is competitive-or-best (within 0.003 of scGen) on 8-9 of 9 metrics on each of 4 datasets. It significantly improves on scGen on eight of nine metrics, favoured in all four datasets on seven; the lone exception, the per-gene log-fold-change correlation, is a tie-or-better everywhere. Against three modern baselines (scPRAM, biolord and CPA) on the same protocol, it beats biolord and CPA on 9/9 metrics and scPRAM on 7/9 (pooled). The affine moment-matching step improves distribution distances (MMD, energy, sliced-Wasserstein), though C2ST stays near 1.0 for all methods. Conclusions: AMM-SimWMag recovers the deep model's advantage over a linear baseline without any neural network. No single deep model is robust across all four biologies, whereas AMM-SimWMag matches or beats them across 4 datasets, CPU-only and reproducibly.