Multi-Modal Foundation · tahoebio

Rhaister

A statistical model predicting cellular perturbation responses across drugs, knockdowns, and cytokines.

What Rhaister is

Rhaister is a perturbation prediction framework developed by Tahoe Therapeutics. Cellular perturbation screens assess how cells respond to compounds, genetic alterations, or signalling molecules across different biological settings, but testing all potential combinations experimentally remains intractable. Rhaister predicts the transcriptional and growth outcomes of unobserved biological context and perturbation pairings.

Rather than modelling individual single-cell transcriptomes directly, the model uses aggregated summary statistics derived from standard single-cell workflows, such as log2 fold changes, expression deltas, and Mann-Whitney U-test p-values. This formulation allows the system to train within seconds and output predictions in milliseconds, matching or exceeding more computationally intensive virtual cell models while approaching empirical half-sample reference limits.

How it works

In the fewshot regime, Rhaister estimates unobserved context and perturbation combinations using a three-stage procedure. First, it calculates an additive baseline using alternating least squares (ALS) to capture global mean, context, and perturbation effects. Second, it employs ridge regression per held-out context to represent unseen perturbations as linear weighted combinations of panel perturbations. Third, a small multilayer perceptron (calibration network) ingests fold change magnitudes, negative log p-values, reference expression, and gene-level statistics to generate calibrated p-values. Three metrics—log2 fold changes, expression deltas, and Mann-Whitney p-values—are predicted jointly, with output p-values formed by an equal blend of ridge regression and calibration network estimates.

For zero-shot prediction on completely unseen cell lines lacking an existing perturbation panel (termed Rhaister-O), the system relies purely on baseline dimethyl sulfoxide (DMSO) pseudobulk expression, applying perturbation- and gene-specific linear coefficients. Training data spans multiple modalities, including the Tahoe-100M small-molecule drug screen (50 cell lines, 384 drugs across 3 doses), the Parse PBMC cytokine screen (12 donors, 18 cell types, 90 cytokines), the Replogle-Nadig CRISPR Perturb-seq screen (4 cell lines, ~2,023 gene knockdowns), and growth-rate datasets such as Emerald Bay and PRISM.

Key facts

Release date: June 2026
Architecture: Alternating least squares (ALS), ridge regression, and a calibration multilayer perceptron
Input data types: Summary statistics (log2 fold change, Mann-Whitney U-test p-values, expression deltas, baseline pseudobulk expression)
Training datasets: Tahoe-100M, Parse PBMC, Replogle-Nadig Perturb-seq, Emerald Bay, and PRISM

What you can use it for

Fewshot small-molecule response prediction: Predicting joint differential expression changes, expression deltas, and significance values for novel drug-cell combinations when a baseline panel has been measured in that context.
Zeroshot drug response forecasting: Estimating post-treatment expression alterations in entirely novel cell lines (Rhaister-O) using only unperturbed baseline (DMSO) pseudobulk expression.
Genetic perturbation modelling: Forecasting transcriptional knock-out consequences across CRISPR Perturb-seq datasets containing thousands of gene targets.
Cell sensitivity and growth rate prediction: Estimating scalar cell growth inhibition and viability metrics across compound screening panels such as PRISM and Emerald Bay.

Run it here

We list this model but cannot run it yet: no pipeline_tag — needs a human.

Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.

Strengths and limits

Good at

Trains in seconds and performs inference in milliseconds by relying on aggregate summary statistics instead of individual single-cell profiles.
Outperforms heavier virtual cell architectures such as STATE while approaching the experimental half-sample reference ceiling.
Jointly predicts directional log2 fold changes, expression deltas, and calibrated statistical significance values.
Supports both fewshot panel-based extrapolation and zeroshot prediction from unperturbed baseline controls.

Watch out for

Performance is markedly lower in CRISPR knockdown settings, achieving a Pearson delta of 0.41 and PR-AUC of 0.08 on Replogle-Nadig data.
Differential expression overlap accuracy drops substantially when operating in zeroshot mode compared to fewshot setups.
Growth sensitivity prediction without supplementary features yields low explanatory power on specific screens, such as Emerald Bay (R² of 0.26).

Inputs and outputs

You provide

Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.

You get back

Mean-pooled representation, one vector per input.

Licence

Check the licence before commercial use · apache-2.0

The apache-2.0 licence carries terms worth reading before you depend on it.

Questions

Can I use Rhaister commercially?

Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.

What inputs are required to run Rhaister?

Rhaister operates on summary-level statistics from differential expression analyses—specifically log2 fold change, expression delta, and Mann-Whitney U-test p-values—or baseline pseudobulk expression profiles for zero-shot runs.

How does Rhaister differ from single-cell foundation models?

Instead of processing single-cell gene expression counts through deep transformer networks, it uses linear combinations, alternating least squares, and a compact calibration MLP over aggregated statistics, enabling orders-of-magnitude faster compute.

What is the difference between Rhaister and Rhaister-O?

Standard Rhaister is a fewshot model that requires a panel of observed perturbations in a target context to infer unobserved pairs. Rhaister-O is a zeroshot variant that predicts responses for novel cell lines using only baseline DMSO control expression.

How to cite it

Svensson, V., Khan, U., Heydari, H., Ubas, A. A., Thomas, N., Merico, D., Goodarzi, H., Yu, J., Alidoust, N., & Gandhi, S. (2026). Back to basics: Observed statistics are sufficient to predict drug responses. bioRxiv, 2026-06. https://doi.org/10.64898/2026.06.09.731197

Written from Rhaister's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.