Arc Institute: Arc's Stack AI Simulates Cell States Without Fine-Tuning

Imagine building AI models for biological simulations that adapt to new drugs or diseases on the fly, without the costly and time-intensive fine-tuning process. For developers and technical buyers in biotech and pharma, Arc Institute's Stack represents a game-changer: a foundation model that unlocks zero-shot predictions of cell behavior, slashing development timelines and enabling scalable, context-aware single-cell analysis directly at inference time.

What Happened

On January 9, 2026, the Arc Institute announced Stack, a groundbreaking single-cell foundation model designed to simulate cell states under novel conditions without requiring fine-tuning. Trained on over 150 million uniformly preprocessed single cells from diverse datasets, Stack employs a novel encoder-decoder architecture with tabular attention mechanisms that facilitate both intra- and inter-cellular information flow. This allows the model to perform in-context learning at inference time, where cells from the query context serve as guiding examples to predict responses to unseen perturbations, such as new drugs or environmental factors. Key capabilities include zero-shot generalization to new biological contexts, generation of hypothetical cell profiles, and efficient embedding of single-cell data for downstream tasks. The release includes open-source code, pre-trained models, and tutorials for integration, making it accessible for researchers to experiment with predictions on custom datasets. The announcement was detailed in a bioRxiv preprint and accompanied by a GitHub repository for implementation [source](https://www.biorxiv.org/content/10.64898/2026.01.09.698608v1), [source](https://github.com/ArcInstitute/stack). Early coverage highlighted its potential to transform perturbational single-cell analysis, with Arc's Twitter post emphasizing the no-fine-tuning breakthrough [source](https://x.com/arcinstitute).

Why This Matters

For developers and engineers in computational biology, Stack's in-context learning paradigm shifts the paradigm from rigid, dataset-specific models to flexible, generalizable AI that mimics how biologists use exemplars to infer outcomes. This reduces computational overhead—fine-tuning large models can require thousands of GPU hours—and enables rapid prototyping of virtual cell experiments, accelerating drug discovery pipelines by simulating untested scenarios. Technical buyers in pharma and biotech stand to benefit from lower barriers to AI adoption: Stack's open-source nature lowers vendor lock-in risks, while its performance on benchmarks like predicting gene expression under perturbations outperforms traditional methods by enabling causal inference from observational data alone. Business implications include faster time-to-insight for R&D teams, potentially cutting costs in high-throughput screening by 50% or more through inference-time adaptations. As AI integrates deeper into life sciences, Stack positions early adopters to lead in precision medicine, where simulating patient-specific cell responses could personalize therapies without retraining models for each cohort [source](https://aihola.com/article/arc-institute-state-virtual-cell-model).

Technical Deep-Dive

Arc Institute's Stack represents a breakthrough in single-cell foundation models, enabling zero-shot simulation of cellular responses to novel perturbations like drugs or diseases without fine-tuning. Trained on vast single-cell RNA sequencing (scRNA-seq) datasets, Stack leverages in-context learning where sets of cells act as prompts to guide predictions in unseen biological contexts.

Architecture Changes and Improvements

Stack introduces a tabular transformer architecture tailored for scRNA-seq data, treating it as a 2D table of cells (rows) and genes (columns). Unlike traditional models that process cells independently, Stack's transformer blocks enable bidirectional information flow: intra-cell (gene-gene relations) and inter-cell (similarity across cells). This captures contextual dependencies, such as how a T cell's state varies in inflamed vs. healthy tissue.

A key innovation is the use of trainable "gene module tokens," which aggregate multiple genes into interpretable biological modules (e.g., pathways), reducing dimensionality and improving efficiency over per-gene modeling. Pre-training on 149 million cells from scBaseCount internalizes broad biological priors across tissues, donors, and states. Post-training on 55 million cells from CellxGene and Parse PBMC datasets teaches in-context learning, allowing Stack to adapt prompts (e.g., drug-treated immune cells) to predict responses in novel cell types like epithelial cells.

Compared to Arc's prior STATE model, which excels in perturbation-specific predictions with dedicated data, Stack generalizes via observational data for hypothesis generation. STATE focuses on expanding experimental perturbations (e.g., new drugs/combinations), while Stack enables broad, zero-shot translation across contexts. Future STATE 2 will integrate both approaches. The model is open-source, with implementation details in PyTorch available at the GitHub repository, facilitating custom extensions.

Example inference pseudocode for perturbation simulation:

import torch
from stack_model import StackModel # Hypothetical import

model = StackModel.load_pretrained("arcinstitute/stack")
prompt_cells = load_scRNA_data("drug_treated_immune_cells.h5ad") # Prompt: observed perturbation
target_cells = load_scRNA_data("untreated_epithelial_cells.h5ad") # Target: novel context
predictions = model.infer_perturbations(prompt_cells, target_cells, perturbation="drug_X")
# Outputs: predicted gene expression shifts for target cells

[source](https://github.com/ArcInstitute/stack)

Benchmark Performance Comparisons

Evaluated on cell-eval, a perturbation prediction benchmark, Stack outperforms baselines like scVI (variational autoencoder) and PCA across tasks including perturbation response, disease classification, and cell-type integration. It achieves superior zero-shot performance, competing with task-specific models without retraining. In 28 of 31 benchmarks, Stack beats prior methods, a rarity in biology where >60% improvement is notable.

Perturb Sapiens, a derived atlas of ~20,000 cell type-tissue-perturbation combinations from Tabula Sapiens, validates predictions against wet-lab data, showing biologically accurate, cell-specific effects (e.g., cytokine responses in epithelial cells). Stack's in-context learning yields predictions matching experimental outcomes better than non-contextual models, enabling in silico screening to reduce $5M+ drug discovery costs.

Developer reactions highlight its paradigm shift: "Zero-shot across tissues/diseases... query biology like a database" (@IterIntellectus on X), praising implications for counterfactual simulations in personalized medicine.

[source](https://www.biorxiv.org/content/10.64898/2026.01.09.698608v1) [source](https://arcinstitute.org/news/foundation-model-stack) [source](https://huggingface.co/datasets/arcinstitute/Perturb-Sapiens)

API Changes and Pricing

Stack lacks a hosted API, emphasizing open-source access via GitHub for research integration. No pricing is specified, as it's freely available for non-commercial use under Arc's license. For enterprise, contact Arc Institute for potential collaborations.

Integration Considerations

Integrate Stack with scRNA-seq pipelines using AnnData/H5AD formats. It requires GPU acceleration (e.g., NVIDIA A100 for large inferences) due to transformer scale. Compatibility with Scanpy/Seurat ecosystems allows seamless embedding in workflows for virtual perturbation atlases. Challenges include handling high-dimensional gene spaces; gene module tokens mitigate this. Early adopters note easy Hugging Face dataset loading for Perturb Sapiens, accelerating drug response modeling without wet-lab iterations.

[source](https://arcinstitute.org/news/foundation-model-stack)