All 7 methods cluster within a 4 pp accuracy band — the gap is smaller than expected.
Primary metric is Macro F1 (handles class imbalance). scGPT pretrained ranks 6th in F1 but achieves 90.7% accuracy with zero task-specific architecture design. The honest finding: on small, clean datasets, pretraining advantage is minimal.
Accuracy & Macro F1 by Method
PBMC3k test set (n = 396), stratified 70/15/15 split, seed = 42. Axis starts at 0.85.
Identical protocol across all methods — no cherry-picking.
From a single fine-tune job to a publication-ready benchmark.
Structured as a benchmark release, not a transient experiment folder.
Strong on B cells and monocytes. CD8 T cells remain the hardest class.
Publication-quality figures from the benchmark paper.
Benchmark lightweight alternatives before deploying a 51M-parameter foundation model.
On PBMC3k, a Deep MLP trained in 14 seconds (Macro F1 = 0.948) outperforms scGPT pretrained (Macro F1 = 0.897). The gap is only 5.1 pp — but the compute difference is orders of magnitude. scGPT's value lies in large, complex, multi-dataset scenarios. This benchmark makes that boundary explicit.