BGPT: Paper Review: Stack: In-Context Learning of Single-Cell Biology

Explore by Goal

Quick Explanation Copied

Stack: context-aware single-cell foundation model

Stack claims that cells-as-context enables inference-time in-context learning: embeddings improve zero-shot performance and prompting (prompt/query cell sets) can generate counterfactual perturbation effects, demonstrated via Perturb Sapiens. Evidence is supported by large-scale training and many evaluation splits, but key uncertainties remain around calibration, protocol/batch sensitivities across perturbation atlases, and how well the “prompt→counterfactual” mechanism generalizes beyond the human immune/epithelial-heavy settings used here.

Primary paper:

Long Explanation

Paper Review (Critical): “Stack: In-Context Learning of Single-Cell Biology”

DOI: 10.64898/2026.01.09.698608 • Paper date (in text): January 09, 2026 • Authors (from text): Dong, Adduri, Gautam, Carpenter, Shah, Ricci-Tam, Kluger, Burke, Roohani

1) What Stack claims (and the core mechanism)

Architecture claim: Stack uses a tabular transformer that mixes intra-cell attention (among gene-module tokens within a cell) and inter-cell attention (among concatenated gene-module tokens across cells in a cell set) to learn representations informed by the rest of the cell set context.
Learning claim: Pre-training uses masked gene reconstruction across all cells in a set with a masking ratio sampled from U(0.1, 0.8), plus latent regularization via sliced Wasserstein distance to enforce a structured latent distribution and avoid shortcuts.
Inference-time “ICL” claim: Post-training aligns Stack to enable cell prompting: provide prompt cells that encode a condition/perturbation/donor and a separate query set whose counterfactual state (gene expression and embeddings) is predicted; generation is performed iteratively with a confidence classifier guiding selective unmasking/replacement.

2) Visuals from the paper’s own scale/data descriptions

The paper provides key dataset sizes; the charts below summarize those numbers (no external estimates).

3) Evidence strength: where the paper is most compelling

3.1 Large-scale pretraining + explicit context modeling

The paper’s emphasis is not just “bigger model,” but a set-level modeling choice: inter-cell attention directly injects information from other cells in the same sample into each cell embedding. This matches their stated goal of overcoming single-cell-level denoising limitations by explicitly leveraging population-scale dependencies at inference.

3.2 Broad evaluation footprint (zero-shot and ICL-style prompting)

The paper reports improvements in zero-shot probing/integration tasks and then extends to prompt-conditioned counterfactual generation and perturbation effect prediction across multiple benchmarks, including a “whole-organism” generated atlas (Perturb Sapiens) that is further validated with external in vitro stimulation datasets.

Skeptical note: “broad evaluation” is persuasive only if metrics are well-calibrated and baselines are carefully matched; the Methods section appears detailed, but some results rely on filtering thresholds and on dataset-specific protocol similarity.

3.3 Mechanistic regularizers aimed at avoiding memorization

Their use of sliced Wasserstein latent regularization is explicitly motivated by identifiable latent variable modeling arguments, and ties to their stated objective of preventing memorization shortcuts while preserving single-cell resolution. Related theoretical background is also cited to identifiable model frameworks.

4) Critical analysis: what could be misleading or incomplete

4.1 “ICL” is operationally prompt-conditioning, not necessarily causal generalization

Stack’s counterfactual generation is evaluated by matching gene-expression and differential-expression metrics in various prompt/query splits. However, these benchmarks test predictive transfer, not a guaranteed causal counterfactual under all perturbation mechanisms. In fact, perturbation responses can include secondary signaling cascades (explicitly acknowledged for TNF-α), meaning “prompt-conditioned generation” may still conflate direct transcriptional effects with downstream propagation effects.

4.2 Calibration and filtering: confidence logits change the story

The Perturb Sapiens analysis includes a classifier-guided filtering step (e.g., restricting to logits below a threshold). This is reasonable for generation quality, but it introduces a selection mechanism that can inflate agreement with validation while reducing coverage. The paper itself shows that unfiltered analyses degrade cell-type specificity, implying that reported performance is partially driven by post-hoc filtering.

4.3 Perturbation atlas protocol/batch effects remain a central uncertainty

The paper evaluates across multiple perturbation datasets with different experimental protocols and uses pseudobulk/DE metrics. Batch effects and dropout imputation differences are well-known to affect differential-expression and correlation-style metrics. Stack’s masking-based training and set-level attention may help, but the remaining sensitivity to cross-assay differences is a known blind spot for any single-cell perturbation predictor. The paper partially addresses this by focusing on splits designed to limit leakage, but protocol differences are still likely to contribute to the remaining error and to overfitting-to-benchmark dynamics.

4.4 “Whole-organism atlas” realism: metrics vs mechanism

Perturb Sapiens is validated using in vitro stimulation datasets for selected cytokines and epithelial lineages. That is scientifically valuable, but it remains selective validation—agreement in DE and correlation metrics does not automatically guarantee correct downstream gene regulatory causality across tissues, especially where prompt/query mismatch or time-dose differences exist (the paper notes stimulation timing/dosage differences in at least one evaluation setup).

5) Practical takeaways for a computational biologist

Set-conditioned embeddings may reduce reliance on dataset-specific fine-tuning by using intra-/inter-cell attention to absorb context at inference.
Prompt/query construction is the real lever: it determines what conditioning information the model sees and therefore what counterfactual mapping it learns.
Evaluation metric choice matters: Pearson-delta vs DE overlap vs direction match can respond differently to batch effects and gene filtering; the paper uses multiple metric families and also adds Jaccard to account for DE set size differences.

6) What would most falsify (or sharply revise) Stack’s headline claims?

Failure of prompt conditioning to generalize: if prompt/query pairs that are disjoint in cell types and donors still fail to improve predictive metrics beyond well-matched baselines (especially under protocol-mismatched perturbations).
Overreliance on confidence filtering: if high agreement only occurs after classifier-based filtering and unfiltered predictions do not show stable biological specificity.
Calibration/uncertainty breakdown: the paper emphasizes scores for confidence but does not show full uncertainty calibration analysis in the provided text; a falsification would be systematic miscalibration across tissues/cell classes.

7) Author reviews to read next (bespoke BGPT links)

Feedback:

Updated: April 05, 2026