2026-07-28
ArchVelo uses archetypal analysis (a classic interpretable convex matrix factorization) to decompose single-cell ATAC chromatin into a few archetypal regulatory programs, coupled into an RNA-velocity kinetic model for trajectory inference; on mouse-brain and human-hematopoiesis benchmarks it beats scVelo, DeepVelo, MultiVelo and three others on accuracy and gene-level latent-time alignment, and decomposes the velocity field into archetype components to nominate driver transcription factors.
Single-cell sequencing takes a gene-expression snapshot of each cell, but cells are dynamic: they differentiate, activate, proliferate. Inferring "who is going where" from a pile of static snapshots is an old, hard problem in genomics. RNA-velocity methods use the ratio of unspliced to spliced RNA and ordinary differential equations to model each gene's transcription kinetics, then assemble a velocity vector per cell. The trouble is that single-cell data is noisy and sparse, so the models are neither robust nor easy to interpret.
Multi-omic sequencing (scATAC+RNA-seq) measures expression and chromatin accessibility in the same cell, which should give more direct regulatory signal. But existing methods like MultiVelo collapse all chromatin peaks of a gene into one value, throwing away the fact that multiple regulatory elements can each respond to different upstream signals. ArchVelo, from the Flatiron Institute and Princeton, targets exactly this.
The core is to hand the chromatin-accessibility data to archetypal analysis. This is a classic unsupervised method in the convex matrix-factorization family (a relative of NMF and PCA): it assumes each cell's profile is a convex combination of a few "extreme" archetypes, each representing a characteristic regulatory program. That both reduces dimensionality and yields interpretable, biologically meaningful bases.
On this low-dimensional representation, ArchVelo models each gene's transcription rate as a linear combination of archetypes (rather than MultiVelo's single aggregated value), then couples it into an extended RNA-velocity ODE system that jointly describes chromatin-accessibility and transcription dynamics. The archetypes are shared across genes, which gives latent-time estimation a common constraint and improves consistency.
On two public multi-omic datasets (mouse embryonic brain and human hematopoietic stem-cell differentiation), the authors compare ArchVelo against scVelo, MultiVelo, DeepVelo, VeloVI, TFVelo and Cell2Fate, covering the classic, deep-network, deep-generative and Bayesian lines:
It also adds a capability the others lack: decomposing the RNA-velocity field into archetype-specific components, and, combined with transcription-factor motif analysis, nominating candidate driver TFs for each component. The authors apply it to CD8 T cells in viral infection, separating functional-differentiation and proliferation trajectories and finding a previously unseen Ccr6-negative to Ccr6-positive path in progenitor exhausted cells.
For single-cell and computational-biology practitioners, ArchVelo is a more accurate and more interpretable multi-omic trajectory tool (open source at github.com/pritykinlab/ArchVelo). For a broader AI audience it is a counterexample worth noting: on a task with clear enough structure, a classic, interpretable, linear method (archetypal analysis plus convex combination plus ridge regression) steadily beats purpose-built deep models (DeepVelo's graph convolution, VeloVI's VAE). Prior structure and an interpretable basis are sometimes worth more than stacking layers.
Most benchmark metrics are distribution-level (scatterplots, log-likelihood, silhouette, CBDir); there is no single clean accuracy number, and the margin over other methods has to be read off the figures. "Beats deep learning" holds on these two datasets under this evaluation setup, and the paper does not sweep systematically across datasets or ground-truth graphs. The number of archetypes K is preset (K=8 is the illustration), with sensitivity analysis only in the supplement. The biological finding still needs wet-lab validation, and a computed trajectory is not a mechanistic proof.