PhysInfer: A physics-constrained neural network for adaptive inference of transcriptional bursting from single-cell RNA-seq.

PhysInfer infers transcriptional bursting from static scRNA-seq by reparameterizing unidentifiable microscopic rates into interpretable burst size and burst frequency, validated on synthetic data.

• Why it matters: Static snapshots suffer from parameter unidentifiability and topology mismatch, limiting reliable recovery of burst dynamics from zero-heavy single-cell observations.
• What they did: The unsupervised framework uses a physics-constrained neural network with a physics-informed loss (stationary-distribution likelihood plus zero-count anchoring), adaptive model selection, and parameter decoupling.
• The result: On a mouse embryonic fibroblast dataset, it reconstructs a transcriptional dynamic manifold across a 2,137-gene secondary BIC set and identifies genes supporting an effective sequential 3-state representation beyond standard 2-state fitting.

AI summary of the abstract · may contain errors · not medical advice · read the paper.