Single-cell perturbation prediction gets a unified base: DCAF ties data pairing, scoring and modeling into one framework

Single-cell perturbation simulation is no longer only a contest of model architecture; training-data pairing, scoring and the prediction model are handled as one pipeline. Earlier models largely learned differences between randomly paired control and perturbed cells, yet current technology measures a cell only once, so learned regulatory patterns absorbed pairing heterogeneity. On multi-cell-type data, DCAF matching yields lower mean KL divergence than random matching, within-cell-type random matching and CINEMA-OT; removing global optimization drops the matched-control ratio below 20% in four of seven datasets.

The prediction model receives both pre- and post-perturbation cell states instead of projecting only control cells into a latent space. Common generative frameworks encode only control cells, whereas DCAF-model uses a joint embedding predictive architecture in which control and perturbed cells pass through encoders whose parameters are linked by an exponential moving average. The encoder is pretrained scGPT finetuned with LoRA, and the predictor is a three-layer transformer whose first block injects perturbation information into control gene embeddings by cross-attention; on Frangieh, H1 hESCs and adult Mouse brain data, accuracy on non-silenced genes exceeds the generative framework.

Scoring looks only at log fold changes of non-silenced genes, making it more sensitive to model error and more stable against zero-inflation noise. Count-based MAE, reference-based correlations and DEG-based metrics each capture different aspects, and non-biological noise from zero inflation disturbs assessment. On synthetic data with known error levels, DCAF-criterion's minimum response is 3.5-fold and 60-fold higher than the best-performing Systema on R² and MSE of log fold change; when noise is injected only into zero-expression genes, its response range is 898-fold and 60-fold lower than Pearson correlation.

Under held-out genes, held-out cell types and cross-region transfer, DCAF-model gives interpretable perturbation-response predictions. Earlier systematic benchmarks showed that complex deep-learning and foundation models do not consistently outperform linear or mean baselines. On Human brain organoids the model is best for unseen genes and for unseen Astrocytes, Excitatory neurons, Inhibitory neurons, NPCs and OPCs; transferring from PFC to MOp in adult Mouse brain, Dgcr8 knockdown predictions recapitulate known cell-type-specific regulatory heterogeneity.

The demonstrated scope covers perturbation data from cell lines, multi-cell-type brain tissue, human brain organoids and adult Mouse brain, scored mainly by log fold change of non-silenced genes with separate MSE and MAE for persistently silenced genes. The intended reader is a researcher working on single-cell perturbation modeling and benchmarking. The source proposes future exploration of sequencing a single cell both before and after perturbation, and of organizing simulation results into experimental design.

The source states that DCAF-criterion's threshold for silenced genes is set manually from each dataset's expression distribution. The cross-cell-type advantage tracks transcriptomic similarity and dispersion between training and test cell types, with higher accuracy when Astrocytes and OPCs are more similar. These conclusions come from a preprint not yet peer reviewed.