scDEBART: Predicting in silico Single-Cell Perturbation Responses via Large-Scale Differential Expression Learning

Single-cell foundation models trained on millions of cells can learn gene expression patterns across diverse contexts. However, for predicting genetic perturbation effects they often underperform simple regression models. We hypothesize two potential limitations: targets defined on dropout-prone absolute expression, and pretraining objectives that reconstruct static co-expression rather than encoding how genes co-regulate under expression changes. We introduce $\textbf{scDEBART}$, a perturbation-specific pretraining framework that predicts log fold-changes (logFC) conditioned on basal expression, thereby learning how gene sets co-vary across expression-change contexts at scale. To obtain reliable estimates of expression change under technical sparsity, we compute logFC from scVI-denoised expression and restrict pretraining to genes with robust detection. Pretrained on 6.28 million expression-change profiles from 66.6 million human cells and fine-tuned on five Perturb-seq datasets, scDEBART achieves mean enrichment factor (EF) of 11.96, 4-7$\times$ higher than scGPT and GEARS (mean EF 1.74-2.99), and 71.4\% top-1 accuracy for reverse perturbation identification compared to near-zero accuracy for prior models. In cross-modal transfer to drug perturbations (SCIPLEX), the model shows dose-dependent improvement in directional alignment (cosine similarity 0.04→0.30) with above-random DEG enrichment (EF 2.91-4.32), suggesting partial transfer of learned regulatory patterns across modalities. Overall, these results indicate that large-scale pretraining on scVI-denoised expression-change profiles provides a useful inductive bias for perturbation prediction.

When a gene in a cell is switched on or off, hundreds of other genes react in response. Predicting these chain reactions computationally would be a powerful tool for discovering new drugs and understanding disease — but current AI models surprisingly struggle to outperform simple statistical methods on this task. We asked why, and identified a key flaw: these models are trained to predict how active each gene is, rather than how much it changes. Gene activity measurements from single cells are extremely noisy — many genes appear silent due to measurement limitations, not biology. When models learn from this noise, their predictions of change are distorted. Our approach, scDEBART, instead learns directly from gene changes across millions of cell comparisons, using a statistical cleaning step to filter out measurement noise before training. Tested across five large genetic perturbation experiments, scDEBART recovered the truly affected genes 4–7 times more accurately than existing models, and could even identify which gene was experimentally altered just by observing the resulting pattern of changes — a task prior models essentially failed.