Single-cell perturbation prediction near experimental reproducibility with GENESIS

Problem Statement
Single-cell perturbation atlases, such as Tahoe-100M and the PBMC cytokine atlas, now measure cellular responses across millions of cells. However, the combinatorial space of biological contexts (cell lines, donors), perturbations (drugs, cytokines), doses, and experimental variables makes exhaustive measurement infeasible. Existing computational approaches attempt to predict perturbation responses by learning mappings from measured control populations to perturbed populations using models like variational autoencoders, diffusion models, or optimal transport. A critical limitation of these methods is that single-cell sequencing is destructive; control and perturbed cells are unpaired samples from condition-level distributions rather than paired measurements of the same cell. Consequently, evaluating these models is difficult because there is no clear standard for how close predictions can get to the agreement achievable between independent experimental repeats. Without a reference for experimental reproducibility, improvements over existing methods do not establish whether predictions approach the reliability of physical measurements.

Methodology: GENESIS
The authors introduce GENESIS (Generative ENErgy-based Single-cell Intervention Simulator), a generative framework that predicts single-cell response distributions directly from the covariates defining an experimental condition (e.g., cell line, drug, dose, plate).

• Architecture: GENESIS employs a lightweight neural network (a multilayer perceptron) that takes learned embeddings of experimental covariates concatenated with random noise as input. It generates individual cell profiles without requiring measured control cells at inference. Control conditions are represented simply by a dedicated value for the perturbation covariate.
• Training Objective: The model is trained using the energy score, a strictly proper scoring rule for distributions. Unlike methods requiring paired cells or fixed-size population targets, the energy score allows learning distributions directly from individual observations. The objective balances two terms: an "attraction" term that pulls generated cells toward the observed cell, and a "repulsion" term that encourages diversity among generated cells to capture cell-to-cell variability.
• Auxiliary Loss: To better capture gene-gene dependencies and improve convergence, a variogram term is added to the loss function, encouraging the model to reproduce the average magnitude of expression differences between pairs of genes within cells.
• Evaluation Framework: The authors establish an experiment-calibrated evaluation framework. Performance is benchmarked against experimental reproducibility (agreement between independent repeats of the same condition) and split-half reproducibility (agreement between random halves of a single experiment). This allows prediction error to be interpreted relative to the inherent noise of the experimental process.

Key Contributions

1. A Generative Framework for Unpaired Data: GENESIS models cellular response distributions directly from covariates, bypassing the need for paired control-perturbed cells, which are experimentally impossible to obtain in destructive sequencing assays.
2. Reproducibility-Centric Benchmarking: The paper introduces a rigorous evaluation standard where model performance is contextualized against the reproducibility of independent experimental repeats, providing a realistic upper bound for prediction accuracy.
3. Data Efficiency: The framework demonstrates that accurate predictions can be achieved with significantly reduced training data, particularly when the reduction strategy prioritizes breadth (more conditions) over depth (more cells per condition).

Results

• Approaching Experimental Reproducibility: On the Tahoe-100M dataset, GENESIS performance reached experimental reproducibility levels in Pearson correlation (0.430 vs. 0.432 for repeats) and energy distance (0.158 vs. 0.154). It fell only 3.4% short in top-100 differentially expressed gene (DEG) overlap. All other tested generative models (PerturbDiff, STATE, CellFlow) and statistical baselines differed significantly from experimental reproducibility.
• Superior Generalization: In held-out biological contexts (cell lines/donors) and held-out perturbations, GENESIS significantly outperformed existing generative models and statistical baselines on both the Tahoe-100M and PBMC cytokine atlases.
• Robustness to Sparse Data: GENESIS trained on only ~5% of the available perturbed training cells (4.6 million of 89.5 million) outperformed competing models trained on the full dataset. Furthermore, sampling fewer cells across more conditions yielded better performance than deeply profiling fewer conditions.
• Biological Fidelity: Generated profiles successfully recapitulated context-dependent transcriptional programs. Using PROGENy pathway analysis, GENESIS captured shared pathway responses (e.g., MAPK suppression) and distinct, context-specific responses (e.g., opposing inflammatory responses in HOP62 vs. Hs 766T cell lines) that were not observed during training.

Significance
The paper claims that GENESIS establishes a framework for predicting cellular responses across subsampled experimental spaces with accuracy approaching experimental reproducibility. By demonstrating that broader, lower-depth measurements combined with generative modeling can expand experimental coverage, the work suggests a complementary strategy for perturbation studies. This approach allows researchers to augment sparsely sampled populations and predict responses in unmeasured conditions while maintaining biologically meaningful variation, effectively extending the reach of perturbation atlases beyond directly measured conditions. The authors note that GENESIS currently generalizes compositionally (requiring each covariate value to have appeared during training) and cannot yet predict responses to completely unseen compounds or cell lines without external information.