A foundation model trained on millions of cells sounds like an obvious upgrade from a conventional single-cell pipeline. It might be. It might also make an analysis slower, harder to interpret, and no better at answering the biological question.
The distinction becomes clear when we separate three decisions that often get bundled together: the data preparation, the model’s operating cost, and the evidence needed to trust the result.
The preprocessing decision is biological
Single-cell data are not interchangeable strings of gene-expression tokens. Tissue handling, cell dissociation, chemistry, ambient RNA, doublets, sequencing depth, mitochondrial signal, and batch structure all influence the data a model sees.
There is no universal preprocessing recipe for scGPT, SATURN, or every other single-cell model. Methods such as SCTransform, highly variable gene selection, batch correction, and mitochondrial filtering have different assumptions and purposes; not every model expects the same input transformation. An aggressive filter can remove a biologically meaningful population. An inappropriate batch correction can erase the very treatment signal an experiment was designed to measure.
The scGPT study, which describes pretraining on more than 33 million cells, is evidence that large-scale representations can support multiple downstream tasks. It is not evidence that the same normalization choices, thresholds, or accuracy will transfer to every tissue, species, or experimental design.
The important output of preprocessing is not just a clean matrix. It is a documented account of what was excluded, transformed, and retained, with checks that the signal under investigation survived.
Compute costs are workload-specific
Training a large model from scratch, fine-tuning a published checkpoint, and running inference are different workloads. GPU memory and runtime depend on model architecture, number of cells and genes, sequence representation, batch size, precision, distributed training design, and available hardware.
That makes the familiar figures of eight A100s for 72 hours or four A100s for 12 hours poor planning defaults without a particular model, code revision, dataset, and benchmark attached. The same goes for saying inference is “free” or guaranteed to fit on one V100. Published weights may be available, but compute, data preparation, and validation still have costs.
A good cost estimate starts with the intended checkpoint and task. Run a representative subset, measure peak memory and throughput, and include repeated experiments, parameter selection, and biological review. The relevant choice might be a conventional pipeline, a pretrained embedding model, task-specific fine-tuning, or a different architecture entirely.
The biological benchmark matters more than scale
Independent evidence argues against assuming that a bigger model wins. In a 2025 Genome Biology evaluation, Kedzierska and colleagues found that zero-shot scGPT and Geneformer representations sometimes underperformed simpler methods on the tasks they tested. That does not invalidate fine-tuning or every potential application. It means a strong baseline is essential.
For cell-type annotation, compare against carefully curated references and simpler classification methods. For integration, test whether batch effects have been reduced without removing genuine biological variation. For perturbation prediction, hold out relevant conditions and compare with simpler mechanistic or statistical baselines.
An impressive embedding plot is not the same as validated recovery of a novel phenotype. Claims about novel cell states deserve orthogonal evidence and appropriate uncertainty, especially where downstream diagnostic or regulated decisions are involved.
Why 50,000 cells is not a decision rule
A suggested threshold of 50,000 cells or five datasets may be a conversational heuristic for scoping work, but there is no primary result here showing those are universal breakpoints. A rare-cell study with fewer cells can be harder than an atlas with millions. Five highly correlated datasets may add less information than two genuinely different conditions.
A more useful decision question is whether the model solves a problem the baseline fails on, with a measurable gain large enough to justify preprocessing, compute, reproducibility, and interpretation costs.
For exploratory research, that might mean improved generalization to an unseen tissue or dataset. For a regulated use, it means validation against the relevant intended use and applicable requirements,not a generic endorsement of a pretrained model.
The moat is the evidence
Foundation models can be powerful biological tools. Their value is conditional, not automatic. The team that defines its question, preserves the biological context, benchmarks against simpler methods, and can reproduce the result is better positioned than one that treats pretraining scale as proof of fitness.
That is the kind of modeling decision McIntosh Consulting helps research teams frame: establish what needs to be learned, compare defensible options, and build a pipeline that makes the result reviewable.