Paper. Gorin, G., Vastola, J. J., Fang, M., & Pachter, L. (2022). Interpretable and tractable models of transcriptional noise for the rational design of single-molecule quantification experiments. Nature Communications, 13, 7620. https://doi.org/10.1038/s41467-022-34857-7 (original), licensed CC BY 4.0. Below is a short summary in our own words; the full text is reproduced under the paper's licence.
Interpretable and tractable models of transcriptional noise for the rational design of single-molecule quantification experiments¶
What this covers¶
How much the fine biophysical detail of transcription — whether a gene's transcription rate varies because of DNA mechanics or because of a fluctuating regulator — leaves a signature in single-cell RNA count data, and how to design models that are detailed enough to test that against real data while still being mathematically solvable. The question sits at the intersection of stochastic modelling of gene expression and the design of single-cell transcriptomic experiments.
The question¶
Single-cell RNA counts vary because of a mix of processes — DNA supercoiling, regulator binding, RNA processing — and it is unclear how much can be learned about which of these processes is actually driving a gene from counts alone. The field's usual descriptive, phenomenological fits (e.g. negative-binomial-like distributions) summarize data without making a specific mechanistic claim, while detailed biophysical models of transcription are mechanistic but usually too complex to analyze and too underdetermined to let two competing hypotheses be told apart from data. The authors wanted models that are simultaneously interpretable (parameters map onto concrete biophysical mechanisms) and tractable (fully analyzable), so that one can work out in advance which experiment would best discriminate between two specific hypotheses about transcription, rather than fitting a generic distribution after the fact.
The approach¶
The paper builds models in which a continuous, randomly fluctuating transcription rate \(K(t)\) drives a standard discrete birth-death process of nascent (unspliced) and mature (spliced) RNA — an SDE (for the rate) coupled to a CME (for the discrete molecule counts). \(K(t)\) is modelled as a one-dimensional mean-reverting stochastic process whose stationary distribution is chosen to be a gamma distribution, which is what makes negative-binomial-like RNA counts fall out naturally. Two specific, biologically motivated choices of driving process are compared: a gamma Ornstein–Uhlenbeck (Γ-OU) process, standing in for transcription rate variation caused by the mechanical/supercoiling state of DNA, and a Cox–Ingersoll–Ross (CIR) process, standing in for variation driven by fluctuations in the copy number of an abundant regulator. Both reduce to simpler, well-known models (the constitutive model and the "mixture" model) in limiting regimes, which lets the new framework be checked against, and shown to unify, results already in the literature. The authors derive exact analytical solutions for the joint nascent/mature RNA distributions, moments and autocorrelation functions of both models (using an SDE–CME isomorphism for the Γ-OU case and a stochastic path-integral method that generalizes further), simulate both processes to validate the analytics, and then use Bayes factors and likelihood-ratio model selection to ask, first in simulation and then on real data, whether the two models can actually be told apart.
What it found¶
Despite their different biological motivations, the Γ-OU and CIR models have identical means, variances, covariances and autocorrelation functions for nascent and mature RNA — so ordinary summary statistics, and any technology that only reports population averages (e.g. bulk RNA-seq), cannot in principle distinguish them. Comparing full joint (nascent + mature) count distributions does distinguish the models: in simulation, around 1,000 cells were enough to separate them over most of parameter space using a log Bayes factor criterion, and using the joint distribution rather than either marginal alone improved distinguishability by roughly an order of magnitude. A parameter-recovery experiment showed the biophysically meaningful parameters (the mean-reversion rate \(\kappa\) and gain \(\theta\)) are fairly identifiable from simulated data when the distribution is overdispersed, though not when it is close to Poisson-like. Applied to real single-cell transcriptomic data from glutamatergic neurons of four mice (31,649 genes after pseudoalignment), a filtering step selected 80 genes for full model fitting; 73 of these converged, and their likelihood-ratio model assignments (Γ-OU-like, CIR-like, or mixture-like) were consistent across the four biological replicates and broadly consistent with a slower, fully Bayesian fit done for a subset of 12 genes. This is presented as a proof of principle that existing single-cell data can be rich enough to support Bayesian discrimination between distinct noise-generating mechanisms, with specific candidate genes identified for each regime.
Limits and context¶
The real-data analysis is explicitly a proof of principle, not a mechanistic claim: the authors note that interpreting the biochemical meaning of specific gene assignments is difficult without also accounting for technical noise and additional downstream RNA processing, which the models presented here omit (though the supplement discusses extensions). Of the 80 filtered genes, 7 were discarded for having an absolute log-likelihood ratio above 150, which the authors took as a sign of failing to converge to a satisfactory optimum rather than as a genuine model comparison. The paper is also explicit that it does not implement the full "rational experiment design" closed loop it motivates in the introduction — it builds the mathematical and computational foundation for that loop (solving the models, showing which kind of data distinguishes them) without running the follow-up wet-lab experiment. The slow-reversion ("mixture") limit is flagged as attractive but potentially biologically implausible, since matching it would require a driving process with an autocorrelation time of weeks given typical mRNA lifetimes. Orthogonal, targeted experiments are described as necessary to confirm whether the models assigned to specific genes actually correspond to the live-cell mechanism, rather than being just the better-fitting of two candidates.
Citation¶
Gorin, G., Vastola, J. J., Fang, M., & Pachter, L. (2022). Interpretable and tractable models of transcriptional noise for the rational design of single-molecule quantification experiments. Nature Communications, 13, 7620. https://doi.org/10.1038/s41467-022-34857-7