Speaker
Description
Deciphering gene regulatory networks is central to understanding how cells orchestrate gene expression programmes. Single-cell RNA sequencing (scRNA-seq) has enabled the investigation of genome-wide regulatory relationships through pairwise gene correlations. Yet, interpreting these correlations is challenging due to confounding factors such as biological sources of covariation or technical noise. Time-resolved scRNA-seq with metabolic RNA labelling provides access to transcription dynamics, offering opportunities to address these challenges. We introduce a modelling framework integrating mechanistic models of gene regulation with machine learning to analyse regulatory relationships in time-resolved scRNA-seq data. We build stochastic models of causal gene regulatory relationships, including direct regulation and co-regulation, capturing key confounders such as extrinsic noise, cell cycle coupling and technical noise, to simulate temporal summary statistics. A neural network supervised by these simulations learns to classify regulatory scenarios, mapping observed correlation patterns to underlying causal models, while leveraging gene-specific priors inferred from the data. Applied to real time-resolved scRNA-seq data, the framework predicts causal gene–gene relationships while quantifying uncertainty. Overall, our approach shows that mechanistically informed machine learning enables interpretable gene regulation inference from single-cell data.
