spVelo (Long et al., Genome Biology 2025)

Long, Liu, Xue & Zhao. Genome Biology 26:239 (2025). Penn State / Yale.

Summary

spVelo extends the veloVI Bayesian deep-generative model to multi-batch spatial transcriptomics: a VAE with two orthogonal encoders — an MLP on (unspliced, spliced) expression and a Graph Attention Network (GAT) on spatial location — fused in latent space, with a Maximum Mean Discrepancy (MMD) penalty across batches for batch-effect correction. It infers per-gene kinetics, a per-cell latent time, and supports uncertainty, trajectory discovery, driver markers, GRN, and temporal cell–cell communication. For the wiki it is a spatial sibling of TopoVelo — but a revealing contrast: spVelo uses spatial info for batch correction and neighbor structure, not to recover physical migration velocity, so it stays ordinal where TopoVelo reaches μm/hour.

Key Claims

  • veloVI-based model. Per-gene transcription/splicing/degradation rates, latent transcriptional state, and per-gene latent times tied via a shared low-dimensional latent variable (veloVI assumptions), fit by variational inference.
  • Spatial + multi-batch. GAT encodes spatial proximity; MMD aligns batch latent spaces — the first to do velocity across multiple spatial batches (prior methods are per-batch).
  • Spatial info helps. Ablation shows integrating spatial location improves velocity + trajectory inference; spVelo’s direction score ranks highest vs scVelo, veloVI, LatentVelo on simulated pancreas + real OSCC.
  • Downstream. Uncertainty, complex (non-linear) trajectory discovery, state driver markers, GRN, and temporal cell–cell communication.

Physical-time grounding (standing lens)

  1. Latent time — ordinal or metric? Ordinal (veloVI-style per-cell latent time); validated by direction/transition scores. No physical units.
  2. Scale degeneracy. Inherited (veloVI-relative); no absolute anchor.
  3. External time anchor. None — and this is the key contrast with TopoVelo: spVelo uses spatial coordinates for batch correction + GAT neighbor structure, not to derive a physical-unit migration velocity. Spatial context ≠ spatial velocity here.
  4. Constant-rate assumptions. Per-gene constant β, γ (veloVI); α per-state. Spatial coupling does not relax rate constancy.

Useful for the spatial-axis story (spatial-velocity): “spatial” methods split into those that merely use space as context/batch structure (spVelo) and those that recover physical migration velocity (TopoVelo). Only the latter touches physical units — and even then by assumption (see physical-time-grounding).

Connections

Contradictions

  • None. Replaces the bookmark-only stub; sharpens the spatial-axis taxonomy on spatial-velocity.