Benchmarking zero-shot single-cell foundation model embeddings for cellular dynamics reconstruction

Abstract Reconstructing cellular trajectories from time-resolved single-cell transcriptomics is fundamental to understanding processes from embryonic development to cancer progression. While single-cell foundation models (scFMs) promise universal biological representations through large-scale pretraining, their capacity to capture the non-linear dynamics governing cell-fate decisions remains uncharacterized. Here we systematically benchmark multiple scFMs across challenging biomedical scenarios involving branching lineages and continuous state transitions. By coupling zero-shot scFM embeddings with dynamic optimal transport, we evaluated their performance against a traditional highly variable gene (HVG) baseline in backtracking progenitor states, interpolating transition intermediates, and extrapolating future fates. We find that zero-shot scFM embeddings underperform the HVG baseline across diverse biological systems, particularly in recovering the distributional complexity of unobserved cells. Mechanistic analysis reveals that current scFM architectures tend to over-compress subtle temporal signals, causing an artificial “linearization” of branched biological structures that may obscure critical divergence points in disease progression. Our findings suggest that while scFMs provide unified cell-state views, the HVG baseline remains more robust for trajectory inference, identifying a fundamental “temporal-compression” bottleneck that must be addressed to develop next-generation, dynamics-aware foundation models.