On the robustness of scRNA-seq foundation models for plant perturbation response prediction under cross-experiment shift

Foundation models for single-cell transcriptomics promise to learn generalizable representations of cellular states. However, recent evidence suggests they often fail to outperform simple machine learning baselines. Furthermore, their ability to generalize across unseen experimental conditions remains poorly understood, particularly in plants, where rigorous evaluation beyond cell type annotation and batch integration is lacking. To address this, we introduce an Arabidopsis thaliana foundation model, scAraFM, and benchmark it across several perturbation conditions under three increasingly challenging protocols: random splits from a single experiment, replicate-based splits, and cross-experiment transfer learning. We found that random splits overestimate performance by up to 30 points relative to cross-experiment evaluations. Across representation strategies, preserving gene identity consistently outperforms the standard pooled embeddings. Moreover, simple baselines using raw reads remain competitive in single-experiment settings, challenging current claims of universal advantage of foundation models. In contrast, under cross-experiment transfer, pretrained representations show added value, particularly with few labelled samples, suggesting that the benefits of foundation models emerge precisely in the regimes that matter for practical deployment. Overall, our results demonstrate that conclusions about foundation models depend critically on the evaluation design, and that preserving per-gene structure aids generalization in downstream tasks, supporting robust predictions across unseen experimental contexts.