Cell2fate (Aivazidis et al., Nat Methods 2025)
Aivazidis, Memi, Kleshchevnikov, Er, Clarke, Stegle & Bayraktar. Nature Methods (2025). Wellcome Sanger / DKFZ / EMBL (GennadyGorin-adjacent biophysics culture).
Summary
Cell2fate is a fully Bayesian RNA-velocity model that linearizes the velocity ODE so a biophysically richer transcription model is solved analytically, and decomposes velocity into modules — giving a biophysical bridge between RNA velocity and dimensionality reduction (modules ≈ factor/PCA loadings). It targets complex / weak / multi-rate transcription that coarse models mishandle, quantifies uncertainty (posterior of cell-specific time as a QC signal), and spatially maps velocity modules onto tissue. For the wiki it is the rigorous, biophysics-forward branch — aligned with the constructive side of velocity-skepticism — and the benchmark’s standout for direction in some settings (though weak overall, see velocity-benchmark-17studies).
Key Claims
- Linearized ODE + modules. du/dt = α_g(t) − β_g u_g; ds/dt = β_g u_g − γ_g s_g, with α_g(t) expanded in integrable basis functions dα_g/dt = Σ_m λ_mi(α_mgi − α_mg). Each basis = a module (ON/OFF, cell-specific timescale T_c); RNA velocity + counts = linear mix of M modules (mixed-membership ≈ factor analysis).
- Biophysical link to dim-reduction. Modules approximate the joint effect of all active regulators within time windows — a mechanistic interpretation of velocity “components.”
- Uncertainty as QC. Fully Bayesian (Pyro, stochastic VI). The posterior CV of cell-specific time → ~0 where dynamics are real, ~1 in steady-state PBMC (correctly flags “no dynamics”).
- Resolves subtle/late dynamics. Captures late granule-neuron maturation and multi-rate-kinetic (MURK) “transcriptional boosts” via sequential module activation, where scVelo / diffusion pseudotime fail; best CBDir across five datasets.
Physical-time grounding (standing lens)
- Latent time — ordinal or metric? Ordinal, but with two refinements: a cell-specific timescale T_c (richer than a single shared clock) and posterior uncertainty on it. Validated by correlation with developmental age / CBDir — relative, not hours.
- Scale degeneracy. Inherited; cell time is relative, no absolute anchor. (The uncertainty quantification is honest about where time is ill-determined — a useful partial answer.)
- External anchor. None — snapshot scRNA-seq (+ spatial mapping, but no labeling/time-series).
- Constant-rate assumptions. α relaxed — time-varying via module basis functions; β, γ gene-specific constants; module params (λ, T_c) shared across genes.
Cell2fate is the biophysics + Bayesian-uncertainty answer to the rigor critique: it doesn’t claim metric time, but it (a) keeps a faithful biophysical ODE and (b) quantifies where the inferred time is trustworthy. For FlowVelo, the uncertainty-as-QC idea is worth borrowing — report where the temporal coordinate is identifiable vs not.
Connections
- Cell2fate — the method entity (upgraded from bookmark-only).
- velocity-skepticism / GennadyGorin — the constructive-biophysics pole it belongs to.
- splicing-kinetics-ode — solved via linearization here.
- manifold-consistent-velocity — “modules ≈ low-dim factors” echoes the manifold view.
- scVelo — the coarse model it improves on; veloVI / RegVelo — Bayesian relatives.
- velocity-benchmark-17studies — Cell2fate is among the 15 (high memory, slower runtime).
- latent-time / physical-time-grounding — ordinal time + cell-specific T_c + uncertainty.
- FlowVelo — borrow the uncertainty-as-QC framing.
Contradictions
- None. Replaces the earlier bookmark-only stub with primary-source detail; confirms its biophysics-forward placement in velocity-skepticism.