Looking forward to engaging with the community at ICML GenBio in Vancouver and discussing how principled benchmarking can advance single-cell trajectory inference!

Grateful to all co-authors and collaborators for their contributions and insight throughout this project. Fabian Theis @fabiantheis.bsky.social, Malte Luecken, Muzzlifah Haniffa @mhaniffa.bsky.social, and Antony Rose.

By providing multiscale, structure-aware benchmarks, scTRAM enables principled model comparison and elevates trajectory representation evaluation from subjective inspection to systematic and reproducible assessment.

This matters because downstream analyses—from developmental biology to disease progression—depend critically on trajectory integrity, which conventional evaluation protocols often overlook due to their qualitative or ad hoc nature.

Our comprehensive experiments reveal distinct representational trade-offs and expose localized embedding failures by dissecting trajectory integrity across biologically relevant contexts.

scTRAM quantitatively evaluates how much single-cell embeddings preserve ground-truth trajectory structures by decomposing fidelity into complementary metrics—topological consistency, manifold continuity, and pseudotime alignment—rather than relying on heuristic visualizations or single surrogates.