Pheno GS speeds up patient cell data comparison with geodesic transport
Pheno-GS: Phenoscape-scale Geodesic Sinkhorn
Machine Learning
Summary
Comparing single-cell data from many patients helps understand differences between them but is hard to do accurately and quickly. The authors designed Pheno-GS, a method that measures distances between such data by capturing the true geometric structure and handling noise and uneven sampling. It works much faster than previous methods and handles large, complex data better. They tested it on simulated data and real biological datasets.
What this means in practice
- •For biomedical data teams: Compute accurate similarity measures between patient single-cell datasets to better capture patient heterogeneity in large cohorts.
- •For machine learning engineers: Implement scalable geodesic transport computations for large, sparse, and uneven datasets in high-dimensional settings.
Authors
Alistair Wilkinson, Christopher J. Tape, Smita Krishnaswamy
Abstract
High-throughput single-cell data is now collected across large patient cohorts. Understanding patient-level heterogeneity from cellular-level data motivates phenoscaping: embedding each single-cell distribution as a "datapoint," with distances given by optimal transport (OT). Computing geometry-aware OT at this scale, between all pairs of patient datasets, remains an open challenge, since existing methods either rely on Euclidean ground metrics that distort manifold structure or fail under sparse, unevenly sampled, or large-scale data. We present \textbf{Pheno-GS} (Phenoscape-scale Geodesic Sinkhorn), which computes accurate, scalable geodesic transport distances under noisy, unbalanced, large-scale settings via three components: ($1$) graph connectivity regularization for well-defined geodesics on sparse/disconnected manifolds; ($2$) an unbalanced OT formulation via KL marginal penalties; and ($3$) a batched matrix algorithm computing all pairwise distances in one heat diffusion (over $200 \times$ faster than Geodesic Sinkhorn for $500$ distributions). We validate Pheno-GS on synthetic benchmarks and a CyTOF perturbation dataset.