PhysInfer: A physics-constrained neural network for adaptive inference of transcriptional bursting from single-cell RNA-seq
Xiaodong Zeng,
Jingwen Zhang,
Shan Li and
Haohua Wang
PLOS Computational Biology, 2026, vol. 22, issue 11, 1-23
Abstract:
Single-cell RNA sequencing (scRNA-seq) has established that gene expression is driven by stochastic transcriptional bursting. However, inferring burst dynamics from static scRNA-seq snapshots is fundamentally constrained by parameter unidentifiability and model topology mismatch. Here, we propose PhysInfer, an unsupervised framework that integrates a physics-constrained neural network with adaptive model selection and parameter decoupling. A physics-informed loss function, combining stationary-distribution likelihood with zero-count anchoring, reparameterizes the optimization from unidentifiable microscopic rates to an interpretable macroscopic burst-quantity representation of burst size and burst frequency, thereby providing an interpretable parameterization for assessing practical recoverability in static data. Validation on synthetic data with known ground-truth parameters confirms high inference accuracy in low-to-medium activation regimes and shows that recovery of BS and productive BF becomes increasingly condition-dependent as the ON-state probability approaches saturation; under different recovery criteria, the condition-specific upper Pon limit for simultaneously recovering BS and productive BF varies with cell number and capture efficiency. Applied to a mouse embryonic fibroblast (MEF) dataset, PhysInfer reconstructs a transcriptional dynamic manifold across the 2,137-gene secondary BIC analysis set, showing cross-gene variation in fitted burst frequency and burst size. A model-specific stationary likelihood combined with zero-count anchoring improves robustness to zero-heavy observations. The adaptive topology module further identifies a distinct subset of genes with stable support for an effective sequential 3-state representation, revealing distributional structure beyond that captured by standard 2-state fitting. PhysInfer thus provides a robust, topology-adaptive, and interpretable solution for characterizing transcriptional kinetic patterns from static single-cell transcriptomic snapshots.Author summary: Single-cell RNA sequencing (scRNA-seq) offers high-resolution static snapshots of gene expression, yet transcription is inherently stochastic and bursting. Inferring such dynamics from static data is fundamentally hindered by parameter unidentifiability, where distinct sets of microscopic rate constants yield indistinguishable observational distributions. To resolve this degeneracy, we developed PhysInfer, an unsupervised framework that employs a physics-constrained neural network to reparameterize the inference target into an interpretable macroscopic burst-quantity representation defined by burst size and burst frequency. This formulation yields an interpretable macroscopic burst-quantity representation of burst size and burst frequency, enables condition-dependent assessment of their practical recoverability from static data, and allows selective separation of resolved model-order assignments from ambiguous cases. Applied to mouse embryonic fibroblast (MEF) data, PhysInfer characterizes cross-gene variation in fitted burst frequency and burst size and identifies a distinct subset of genes with stable support for an effective sequential 3-state representation beyond that captured by the conventional 2-state model. Collectively, PhysInfer provides a topology-adaptive computational framework with empirically characterized practical-recovery regimes, supporting cautious interpretation of transcriptional kinetic patterns from static transcriptomic snapshots.
Date: 2026
References: Add references at CitEc
Citations:
Downloads: (external link)
https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014904 (text/html)
https://journals.plos.org/ploscompbiol/article/fil ... 14904&type=printable (application/pdf)
Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.
Export reference: BibTeX
RIS (EndNote, ProCite, RefMan)
HTML/Text
Persistent link: https://EconPapers.repec.org/RePEc:plo:pcbi00:1014904
DOI: 10.1371/journal.pcbi.1014904
Access Statistics for this article
More articles in PLOS Computational Biology from Public Library of Science
Bibliographic data for series maintained by ploscompbiol ().