Single-cell perturbation prediction moves beyond variable genes: scBIG++ generates all retained genes and reaches 0.8521 response correlation on Norman combinations

The output range of perturbation response prediction expands from highly variable genes to all protein-coding genes retained after preprocessing in each experimental setting, spanning 6,539 to 19,451 genes. Earlier perturbation-prediction settings mostly focused on highly variable genes, so expression changes in excluded genes were unavailable for downstream analysis of coordinated pathway-member changes. Six genetic and chemical settings derived from Norman, Replogle and ComboSciPlex each use their complete post-preprocessing gene universe, so benchmark comparisons and program analyses evaluate predictions in the same gene space.

A gene's contribution within a shared window varies continuously with cellular expression state, while the window itself stays comparable across cells. In the earlier scBIG, each gene kept the same module assignment across all cells, even though cells with similar identities express programs such as G1/S and G2/M to different degrees. On the Norman additive setting, removing Dynamic SoftSub lowers full-gene response correlation from 0.8521 to 0.8157 and raises MSE from 0.0015 to 0.001940; shuffling the learned assignments lowers it to 0.8271.

Each control cell's representation can draw on information from other cells in the same source population at corresponding gene windows, and a pooled population summary conditions the latent flow together with the perturbation representation. Encoding each cell in isolation prevents other observed control cells from informing its representation. Raising the source set from one cell to 32 raises full-gene response correlation by 33.2% and roughly halves MSE; the default set size is chosen by validation response correlation, and at a set size of 128 cells the full-gene response correlation is 0.8521 with MSE 0.0015.

In the held-out-compound setting the perturbation description becomes a Morgan fingerprint while the full-gene encoder, population conditioning and flow interface are retained. Chemical and genetic perturbations share the same conditioning interface, without a separate drug-specific prediction architecture. On ComboSciPlex, relative to PerturbDiff among the displayed methods, correlation on the top 20 response genes improves by 8.8% and MSE falls by 17.9%.

Evaluation covers six settings derived from Norman, Replogle and ComboSciPlex over 6,539 to 19,451 genes, including unseen gene combinations, held-out targets, cross-cell-line transfer and held-out compounds; benchmark entries are single-seed estimates, and the cross-cell-line conclusion is specific to the displayed K562 transfer protocol.

The single-seed protocol leaves uncertainty about small ranking differences; across all 32 test combinations and 13 programs the mean MAE is 0.0546 for scBIG++ versus 0.0508 for simple addition, so case-level gains do not amount to a uniform advantage over addition; the MHC-II antigen-presentation panel's MAE rises from 0.0046 to 0.0088.