VCBench: A Multi-Dimensional Benchmark for Single-Cell Foundation Models
Single-cell foundation models are increasingly positioned as virtual cells, yet their capabilities are assessed by fragmented, largely single-task benchmarks that obscure where these models improve on simple baselines. VCBench addresses this by synthesizing four independent virtual-cell frameworks into seven capability dimensions: perturbation response prediction, cross-species universality, gene regulatory network (GRN) inference, modality integration, temporal dynamics, multi-scale integration, and in silico experimentation. Each dimension is assessed for operational testability under current architectures and datasets: five admit direct or proxy evaluation, while multi-scale integration and in silico experimentation are structurally untestable as end-to-end tasks. We evaluate five foundation models (Geneformer, scGPT, UCE, TranscriptFormer, Arc State) against pre-registered linear and nearest-neighbor baselines across the five testable dimensions, and report three findings. First, the baselines match or exceed every foundation model on four of the five scored dimensions, replicating the reported competitiveness of linear baselines on perturbation prediction and extending it to cross-species transfer, GRN inference, and temporal ordering. Second, TranscriptFormer alone exceeds the strongest baseline on cross-modal RNA-to-protein prediction (53% Pearson improvement, with a documented contamination caveat) and is the only model to reach Level 2 in the pre-registered Virtual Cell (VC) Level rubric; the architectural choice behind this advantage simultaneously causes a spectral collapse that destroys its temporal-ordering performance, a tradeoff invisible to single-task benchmarks. Third, no foundation model publishes a complete cell-level training manifest, leaving data contamination undetectable to users. Alongside the benchmark, VCBench releases a Contamination Reporting Schema and contributes two further methodological tools: a common-label-set protocol that controls for class-count confounds in cross-species transfer, and a spread-error correlation probe for epistemic calibration.