Rhaister

Rhaister is a perturbation-response predictor that deliberately steps back from the "virtual cell" paradigm. Instead of simulating single-cell responses to a drug or genetic perturbation, it operates directly on the summary statistics that standard single-cell analysis pipelines already produce: per-gene log2 fold change, Mann-Whitney U-test p-values, and expression deltas. From these aggregated observations, Rhaister predicts how a cellular context will respond to perturbations that were never measured in that context.

The central claim of the work is that these observed statistics are sufficient to predict drug responses — that much of the accuracy attributed to large neural virtual-cell models can be recovered by carefully modeling the summary statistics themselves. By measuring only a handful of perturbations in a new biological context, Rhaister infers the unmeasured perturbations by learning how response patterns vary across a reference panel of contexts. The model trains in seconds and predicts in milliseconds, making it compatible with rapid, iterative, and agentic screening workflows.

Rhaister was introduced in a June 2026 bioRxiv preprint, "Back to basics: Observed statistics are sufficient to predict drug responses," by Valentine Svensson, Umair Khan, Hamed Heydari, Airol A. Ubas, Nicole Thomas, Daniele Merico, Hani Goodarzi, John Yu, Nima Alidoust, and Shreshth Gandhi at Tahoe Therapeutics. It sits alongside the company's Tahoe-100M perturbation atlas and the Tahoe-x1 foundation models, but takes an intentionally minimal, statistics-first approach to the same prediction problem.

#Key Features

• Summary-statistic inputs: Rhaister consumes the outputs of routine differential-expression pipelines — log2 fold change, Mann-Whitney p-values, and expression deltas — rather than raw single-cell counts, sidestepping the cost of full single-cell simulation.
• Few-shot perturbation prediction: The few-shot model combines additive alternating least squares for baseline context and perturbation effects with ridge regression that expresses an unseen perturbation as a weighted combination of measured panel perturbations.
• Zero-shot Rhaister-O: A companion variant predicts responses in entirely new cell contexts from baseline pseudobulk expression alone — presented as the first zero-shot model for this task.
• Calibrated significance: A small MLP calibration network maps statistical features to calibrated p-values, sharpening differential-expression calls.
• Extreme efficiency: Training completes in seconds and inference in milliseconds, in contrast to compute-intensive virtual-cell models.

#Technical Details

Rhaister is a compact, statistics-first model rather than a deep network. Its few-shot form couples additive alternating least squares (separating baseline context effects from perturbation effects) with drug ridge regression, then a lightweight MLP for p-value calibration; the zero-shot Rhaister-O uses the form y(c*, p, g) = β(p,g) + γ(p,g) · x(c*,g), where x is baseline pseudobulk expression. It is trained and evaluated across Tahoe-100M (50 cell lines, 384 drugs at 3 doses), Parse PBMC (12 donors × 18 cell types, 90 cytokines), Replogle-Nadig (4 cell lines, ~2,023 CRISPR knockdowns), and phenotypic growth-rate data from Emerald Bay and PRISM. On few-shot transcriptional prediction it reaches a Pearson delta of 0.87 and Spearman LFC of 0.81 on Tahoe-100M, with lower but positive scores on the harder Parse PBMC and Replogle-Nadig benchmarks; zero-shot Rhaister-O reaches Pearson delta 0.63 on Tahoe-100M, and drug-sensitivity prediction reaches R² of 0.87 on PRISM. Across these benchmarks the approach matches or surpasses the leading virtual-cell model, STATE, and its accuracy scales with the diversity of the reference data.

#Applications

Rhaister targets computational biologists and drug-discovery teams who run perturbation screens and want to extrapolate beyond the conditions they can afford to measure. Given a few observed perturbations in a new cell line, tissue, or donor, it predicts the response to the rest of a compound or genetic panel, prioritizing candidates for follow-up. Its speed and reliance on standard pipeline outputs make it well suited to interactive analysis and automated, agent-driven screening loops, where a heavyweight virtual-cell model would be impractical.

#Impact

Rhaister is a pointed counterargument in the virtual-cell debate: it shows that a simple model built on observed summary statistics can match or exceed far more expensive neural simulators on the field's standard transcriptional benchmarks, sometimes saturating the metrics on data-rich atlases like Tahoe-100M. By releasing the code and datasets under Apache-2.0, Tahoe Therapeutics offers both a strong, cheap baseline that future perturbation models must beat and a reminder that careful use of existing statistics can rival end-to-end learning. As a 2026 preprint, its results await peer review and independent benchmarking, and its accuracy depends on having a sufficiently diverse reference panel.