Beyond Visual Inspection: Principled Benchmarking of Single-Cell Trajectory Representations with scTRAM

Single-cell representation learning aims to construct low-dimensional embeddings that preserve biologically salient manifold structures. Yet current evaluation protocols remain inadequate for assessing trajectory fidelity—essential for studying cell differentiation, immune responses, disease progression and perturbation—relying on heuristic visualizations or scalar surrogates that overlook multiscale trajectory geometry. To address this gap, we present scTRAM—single-cell trajectory representation metrics—a principled benchmarking framework that quantitatively evaluates how well single-cell embeddings retain ground-truth trajectories across complementary failure modes, from local neighborhood scrambling to global branch mis-ordering. Through our experiments, we show that scTRAM reveals model-specific trade-offs in preserving distinct aspects of trajectory structure, with edge-specific decomposition exposing localized representational biases at high resolution. Systematic validation across diverse biological contexts confirms that the metric suite is sensitive to genuine structural properties rather than statistical artifacts. By shifting from qualitative, ad-hoc judgments to quantitative, principled approach, scTRAM enables rigorous comparison of embedding methods and informs downstream analyses where trajectory integrity is essential for biological interpretation.