Maturation Delays and MicroRNA Degradation Fields in Spatial RNA Velocity

When a gene switches on, a cell does not instantly fill with mature RNA. New transcripts first appear in an unspliced form — still carrying the non-coding segments that are cut out later — and only then convert into spliced, functional message. RNA velocity methods read the ratio of these two pools as a compass needle: genes with more unspliced RNA than expected are heading up, genes with less are heading down. Applied to spatially resolved tissue slices, this promises a movie of differentiation and disease progression from what is really a static photograph. But the compass has two systematic blind spots, and both were documented in bulk or single-cell settings years before spatial velocity tools shipped.

First, the conversion is not instantaneous. Honkela and colleagues paired RNA polymerase binding with mature RNA over a stimulation time course and found that for 11% of active genes (204 of 1,814) mature RNA lags more than 20 minutes behind transcription, with some lags exceeding an hour, because splicing happens after transcription completes in short genes rather than during it (Genome-wide modeling of transcription kinetics reveals patterns of RNA production delays). Second, degradation is not uniform across a tissue. MicroRNAs — short regulatory RNAs that bind the tail of a target message and speed its destruction — are expressed in spatial gradients, yet every spatial velocity model we found assumes a single degradation rate per gene everywhere. This proposal builds a spatial RNA velocity model that corrects both blind spots, and then closes the loop with the one measurement that can physically check the corrected numbers: the dual-luciferase reporter assay, in which a suspected regulatory sequence is fused to a light-emitting gene so that direct silencing shows up as reduced luminescence, with a second, untargeted light-emitting gene in the same well canceling out transfection noise.

Core Claim

The claim is that velocity fields computed on spatial tissue are biased in two predictable directions: genes with long maturation delays get misread as bursting transcriptionally when their unspliced pool is really just slow to process, and genes sitting in microRNA-rich neighborhoods get misread as already downregulating when their spliced pool is really just being cleared faster. Modeling an explicit per-gene maturation delay $\Delta_g$ and a degradation rate that varies with local microRNA abundance will fix both biases, and — crucially — the inferred microRNA-target strengths can be calibrated against luciferase measurements, provided the assay's own protein-maturation lag is itself modeled rather than ignored. No existing work does any part of this. The spatial velocity tools spVelo (spVelo: RNA velocity inference for multi-batch spatial transcriptomics data) and STT (Spatial transition tensor of single cells) assume instantaneous transcription-to-unspliced appearance and treat degradation as exogenous; cell2fate (Cell2fate infers RNA velocity modules to improve cell fate prediction) states its constant degradation rate as an explicit limitation; DeepKINET (A deep generative model for estimating single-cell RNA splicing and degradation rates) infers state-dependent degradation but has no spatial coordinates and no maturation delay; and the luciferase literature (Serum miR-381-3p: Diagnostic Role and Mechanisms in Liver Transplant Ischemia-reperfusion Injury; miR-944 inhibits malignant progression of bladder cancer through ATIC/AKT/FOXO3 A axis mediated by SHMT1) runs purely as an isolated bulk end-point assay, never as a calibration instrument for a dynamical model.

Research Program

The core object is a Bayesian generative model — call it LagVelo — built in the lineage of veloVI (Deep generative modeling of transcriptional dynamics for RNA velocity analysis in single cells) and cell2fate, but with a spatial graph attention component as in spVelo and, critically, a delayed and spatially coupled rate structure. For gene $g$ at spatial position $\mathbf{x}_c$, with $u$ and $s$ the unspliced and spliced pools, $\alpha$ the transcription rate, and $\beta$ the splicing rate:

$$\frac{du_g}{dt} = \alpha_g(t-\Delta_g) - \beta_g u_g, \qquad \frac{ds_g}{dt} = \beta_g u_g - \Big(\gamma_{g,0} + \sum_k w_{gk}\,\mu_k(\mathbf{x}_c)\Big) s_g$$

Here $\Delta_g \ge 0$ is the maturation delay between transcription and appearance of detectable unspliced RNA, $\mu_k(\mathbf{x}_c)$ is the smoothed local abundance of microRNA $k$ around position $c$, and $w_{gk}$ is a per-pair targeting weight — the quantity we ultimately want to calibrate against luciferase data, since it predicts how much faster message $g$ decays where microRNA $k$ is present. Both new ingredients are exactly what existing frameworks lack: Honkela's bulk ODE had $\Delta$ but no cells, no space, and a constant decay; DeepVelo (DeepVelo: deep learning extends RNA velocity to multi-lineage systems with cell-specific kinetics) and DeepKINET relaxed constant rates but have neither $\Delta$ nor coordinates.

Identifiability is the honest central risk: $\Delta_g$ and $\beta_g$ can trade off, and the splicing/degradation co-dependence problem already noted for joint generative models gets worse, not better, with a third rate. We handle this three ways rather than by hand-waving. Genes get informative priors on $\Delta_g$ keyed to gene length and intron-retention status, exploiting Honkela's finding that short genes with late intron retention carry the long delays while long genes splice co-transcriptionally. The delay term is admitted per gene only when a model-comparison criterion (held-out likelihood, not training fit) prefers it, so the 11%-of-genes regime is discovered, not assumed. And the whole rate structure is first tested on simulations where the truth is known before any tissue claim is made. Velorama (Causal gene regulatory analysis with RNA velocity reveals an interplay between slow and fast transcription factors) showed that transcriptomic lags between regulator and target are recoverable in dissociated data — it distinguished slow chromatin-remodeling transcription factors from fast ones — which motivates our delay parameter; but Velorama has no spatial coordinates and no physical validation, which is precisely where we extend it.

The luciferase arm is where the proposal becomes falsifiable rather than another velocity variant. Standard practice, as in the miR-381-3p/JAK2 hepatic ischemia study and the miR-944/SHMT1 bladder cancer study, transfects a wild-type target 3′ UTR fused to firefly luciferase alongside a mutated seed-matched construct; rescue of luminescence in the mutant proves direct, sequence-specific silencing. We keep that logic but change what is measured and how it is read. Instead of one end-point ratio, we take time-resolved luminescence after perturbing the candidate microRNA, and we fit a small reporter ODE — light output $L$ rises from translation of the reporter message after a delay $\tau$ and falls with the enzyme's own turnover $\delta_L$:

$$\frac{dL}{dt} = \kappa\, s^{\text{rep}}(t-\tau) - \delta_L L$$

This step is forced by the reporter problem: work on yeast gene repression (Altered expression response upon repeated gene repression in single yeast cells) shows that stable reporter protein masks the true timing of transcriptional shutdown, and cell-cycle comparative transcriptomics (Comparative cell cycle transcriptomics reveals synchronization of developmental transcription factor networks in cancer cells) shows protein peaks lag transcript peaks by one to two cell-cycle phases. Raw luminescence is a low-pass-filtered version of the RNA dynamics we are trying to check; the ODE inverts that filter, and destabilized luciferase variants shrink the filter's time constant. The output is a measured decay-rate fold-change per microRNA-target pair, directly comparable to the model's predicted $w_{gk}$.

Proposed Experiments

Three stages. First, synthetic: we extend the SERGIO-style simulated differentiation landscapes used by Velorama with spatial microRNA fields and known per-gene delays, so ground truth exists for $\Delta_g$, the $\gamma_g(\mathbf{x})$ field, and the true velocity vectors. Second, real data: Stereo-seq mouse brain slices (the benchmark tissue in both spVelo and STT), the chicken heart and epithelial-mesenchymal-transition datasets from STT, and the human fetal cortex sections that cell2fate projected onto space with cell2location. Direction quality is scored by Cross-Boundary Directional Correctness, the metric on which cell2fate reports beating ten other velocity methods across five datasets (dentate gyrus, pancreas, erythroid, human and mouse bone marrow), and by cosine similarity to ground-truth velocities on simulation, as in STT's benchmarks. Because dissociated ground truth for kinetics exists, we also run LagVelo on the scEU-seq cell-cycle and scNT-seq hematopoietic metabolic-labeling datasets from DeepKINET, where synthesis, splicing, and decay rates were directly measured, and ask whether $\Delta_g$ and $\gamma$ recovery matches a method designed for exactly that regime. Third, the calibration experiment: for the top-ranked predicted microRNA-target edges with spatially varying degradation, we build a paired wild-type/seed-mutant 3′ UTR luciferase panel in the matching cell system, following the assay design of the miR-944/SHMT1 study in bladder cancer cells, and test whether the model-predicted repression magnitude tracks the ODE-deconvolved measured decay-rate change across the panel, with mutant rescue as the specificity control.

Baselines

The naive reference point is diffusion pseudotime computed on spliced counts alone, which orders cells along a trajectory but has no inherent direction; as Hie and colleagues note in their methods review (Computational Methods for Single-Cell RNA Sequencing), pseudotime cannot distinguish forward differentiation from reversal without external anchors, and Velorama quantified the cost of that ambiguity — incorporating RNA velocity improved causal network inference by 1.8–3× in area under the precision-recall curve compared with pseudotime-based inference. Beating pseudotime is necessary but not the bar. The standard constant-rate baseline is the scVelo expectation-maximization model, which assumes fixed transcription, splicing, and degradation rates, no delay, and no space; veloVI reports a 5× runtime speedup over the scVelo EM model on a 20,000-cell dataset, a computational rather than accuracy figure, so we treat it as the field's default rather than its best. The strongest competitor on spatial data is spVelo, which combines a graph attention network over spatial proximity with a variational autoencoder and multi-batch correction; it reports consistently outperforming scVelo, veloVI, and LatentVelo on velocity confidence and direction scores across simulated pancreas, oral squamous-cell carcinoma, and Stereo-seq mouse brain data, though no reported numerical value was found on Lacuna for those comparisons. We will run spVelo on identical slices with identical preprocessing and an equal hyperparameter-search budget, and beating it means three things: higher direction and confidence scores on its own benchmarks, unbiased delay recovery on synthetic data with ground truth (a task spVelo cannot attempt), and luciferase-confirmed degradation predictions. For rate estimation on dissociated data, DeepKINET is the strongest baseline, reporting higher correlation with metabolic-labeling ground truth than cellDancer and DeepVelo, again with no reported numerical value found on Lacuna; we run it on the same scEU-seq and scNT-seq data with the same tuning budget, and our claim over it is narrower — matching its kinetic-rate recovery while additionally showing that adding spatial coordinates changes the degradation estimates in ways that luciferase assays confirm. Finally, two ablations isolate the proposal's ingredients: LagVelo with $\Delta_g$ forced to zero (delay removed, spatial degradation kept), and LagVelo with $w_{gk}$ forced to zero (uniform degradation, delay kept), plus a validation ablation that reads luciferase as raw end-point ratios instead of through the reporter ODE.

Why not just…?

Why not just use pseudotime and skip velocity entirely? Spatial data are noisy and unspliced reads are sparse, so the simplest trajectory story is tempting. But pseudotime has no arrow of time, and the cost has been measured: Velorama found that replacing pseudotime with velocity-based transitions improved causal inference AUPRC by 1.8–3×. Our target quantity — a per-gene delay and a spatially varying decay rate — is not even expressible in a pseudotime ordering, so the simpler tool answers a different question.

Why not just let the existing splicing rate absorb the delay? Standard velocity already has a splicing parameter $\beta$, so a reviewer will call $\Delta_g$ over-parameterization. This is a real threat and the ablation exists precisely to answer it: if setting $\Delta_g=0$ fits held-out data equally well, the simpler model wins and we will report that. But there is direct evidence the single-rate form fails — Honkela's time-course analysis found 204 genes whose mature RNA lags transcription by more than 20 minutes in a way a single instantaneous conversion rate could not reproduce, and Schweikert's kinetic CRAC work (Kinetic CRAC uncovers a role for Nab3 in determining gene expression profiles during stress) showed steady-state RNA lags instantaneous transcription-rate shifts during stress. The honest caveat is that $\Delta_g$ matters only for a minority of genes, which is why we admit it per gene by model comparison rather than globally.

Why not just use metabolic labeling instead of splicing ratios? Pulse-chase datasets like scEU-seq and scNT-seq measure turnover directly and would beat any inference given fresh, dissociable cells. But they require live chemical incubation, which is impossible for banked clinical tissue and destroys the in-situ architecture that spatial platforms preserve — the tumor-immune hubs that Starfysh (Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor-immune hubs) recovers simply do not survive as a labeling prep. Our design concedes the point where it is strong: we use metabolic labeling as ground truth on dissociated data to check the model's rate recovery, and claim added value only where spatial context exists.

Why not just use TargetScan-style sequence prediction and skip luciferase? Sequence complementarity predicts candidate microRNA targets cheaply, but it cannot prove in-cell repression or its magnitude — an observed expression drop may be an indirect effect through an upstream activator, which is exactly why the clinical microRNA studies pair predictions with seed-mutant reporter rescue. TargetScan also offers no number comparable to a decay-rate fold-change, so it cannot calibrate $w_{gk}$; it will serve as a prior on candidate edges, not a substitute for the assay.

Decisive Checks

The proposal stands or falls on three results. On simulation, LagVelo must recover known delays and spatial degradation fields with low bias, and removing either ingredient must produce the predicted direction errors — misattributed bursts without $\Delta$, false downregulation without spatial $\gamma$ — while the ablations quantify how much each fix actually buys. On real spatial tissue, velocity direction scores must not degrade relative to spVelo; a method that fixes bias but loses directional accuracy overall has traded one error for another. And on the luciferase panel, the measured decay-rate fold-changes must correlate with predicted $w_{gk}$ across constructs, with full rescue in seed-mutants as the specificity gate — the first quantitative bridge between a spatial dynamical transcriptome model and the bench assay that has, until now, lived entirely outside it.