scMTNI: Leveraging cellular trajectory and context to infer dynamic GRNs from single-cell multi-omics data

Abstract: Transcriptional gene regulatory networks (GRNs) depict the directed relationships between regulators and target genes, determining gene expression patterns in a cell-type-specific manner. Single-cell multi-omics technologies, such as single-cell RNA sequencing (scRNA-seq) and single-cell Assay for Transposase-Accessible Chromatin using sequencing (scATAC-seq), enable high-resolution measurement of cell-type-specific gene expression and regulation in an unprecedented way. However, tools for inferring cell-type-specific GRNs and modeling their dynamics remain scarce. To facilitate the inference and analysis of cell-type-specific GRNs in contexts such as cellular development or disease progression, where cell lineage structure and dynamics are important, we developed a multi-task learning framework, single-cell Multi-Task Network Inference (scMTNI). scMTNI and its associated network analyses tools offer a comprehensive package to define cell-type-specific GRNs and examine their dynamics. This book chapter describes the scMTNI tool and demonstrates its application to an existing cellular reprogramming single cell multi-modal dataset to infer cell-type-specific GRNs and identify key regulators of cellular fate transitions during cellular reprogramming.

• The paper presents a novel framework using probabilistic graphical models and multi-task learning to infer dynamic, cell-type-specific GRNs.
• It integrates scRNA-seq and scATAC-seq with cellular lineage priors, employing stability selection and k-means clustering to robustly recover regulatory edges.
• Application to mouse reprogramming data validates the method’s accuracy in identifying both canonical and transient regulators, supported by metrics like a 0.85 matching score.

Detailed Analysis of scMTNI: Dynamic GRN Inference from Single-Cell Multi-Omics Data

Introduction and Motivation

The elucidation of gene regulatory networks (GRNs) at cell type and lineage resolution is central to understanding state transitions in developmental biology and disease. scMTNI (single-cell Multi-Task Network Inference) is designed to infer dynamic, cell-type- or state-specific GRNs by leveraging single-cell multi-omics datasets, integrating scRNA-seq and scATAC-seq, and explicitly modeling known or inferred cellular lineages. Compared to traditional GRN inference methods that ignore either lineage context or the regulatory prior encoded by chromatin accessibility, scMTNI offers a probabilistic and multi-task learning framework that captures both the commonality and divergence of GRNs across related cell states.

Methodological Framework

Probabilistic Graphical Approach

scMTNI formalizes each cell cluster or state as a distinct "task" within a multi-task learning paradigm. Each task is represented with a dependency network, modeling gene expression as a probabilistic function of regulators, informed by both motif presence and empirical chromatin accessibility. Two principal priors guide inference:

1. Cell-Type-Specific Motif Prior: The presence and strength of motif binding sites, stratified by chromatin accessibility, are incorporated via a logistic prior favoring edges supported by accessible TF motifs. The influence of this prior is tunable by the hyperparameter $\beta_1$ (motif penalty), which acts additively with a sparsity parameter $\beta_0$.
2. Cell Lineage Tree Prior: The lineage tree of cellular states (either experimentally well-defined or inferred via, e.g., PAGA MST from expression data) imposes structural constraints on GRN evolution. Three parameters—root edge rate ($p_r$), edge gain rate ($p_g$), and edge maintenance rate ($p_m$)—define the prior probabilities for edge initialization, acquisition, and inheritance across descendant cell types. This prior enforces gradual network divergence and favors network inheritance along lineage branches.

Data Integration and Input Requirements

scMTNI is implemented as a C++ package taking as input a normalized scRNA-seq matrix, cell-cluster assignments, the cell lineage topology, and optionally, cell-type-resolved TF/gene regulatory priors from accessible motif instances (derived from scATAC-seq and motif databases such as CIS-BP). Gene, regulator, and motif mapping files are formatted with cluster-specific gene IDs, enabling both intra- and inter-cluster analyses.

Inference Strategy and Stability Selection

Inference proceeds via blockwise conditional Gaussian modeling of regulator-target dependencies. For robust edge calls and confidence scoring, scMTNI deploys stability selection over multiple input subsamples. Edge confidence is quantified by the proportion of subsamples in which an edge is recovered, supporting downstream filtering either by top $k$ edges or confidence thresholds.

Application to Mouse Cellular Reprogramming

scMTNI was applied to a comprehensive mouse reprogramming dataset spanning six developmental stages (MEF, FBS-Day3/6/9/12, mESC) and comprised of eight cell clusters after integration of scRNA-seq and scATAC-seq via LIGER. Motif prior networks were constructed using ArchR-processed cluster-specific ATAC peaks. The cell lineage topology was inferred via PAGA, resulting in a minimum spanning tree connecting the eight clusters, which served as a guide for lineage-informed GRN inference.

Output and Network Consensus

For each cell cluster, scMTNI produces a consensus GRN capturing the set of regulator-target edges most consistently supported across stability subsamples. Two standard output modes are supported: top 1,000 edges by confidence, or all edges with confidence $\geq 0.8$, facilitating downstream dynamic network analyses.

Dynamic GRN Analysis

k-means Edge Clustering

Using the top high-confidence regulatory edges, k-means clustering (evaluated across a sweep of $k$ using silhouette index) was used to partition edge dynamics across clusters. Each edge’s confidence across clusters forms the basis for clustering, enabling identification of both static and highly dynamic edges/regulators. Noteworthy findings include the correct prioritization of well-established pluripotency regulators within mESC clusters (e.g. Esrrb, Klf2, Zfp42) and lineage-transient regulators such as Tbxt (identifying a previously underappreciated, mixed-relapsed state in cellular reprogramming trajectories).

LDA Topic Modeling

Latent Dirichlet Allocation was adapted to model regulator-target subnetworks as “topics”, with regulators as documents and targets as words. This abstraction reflects families of regulatory programs co-active across subsets of cell states. LDA effectively identified regulators with extensive network rewiring (e.g., Ybx1), supporting a mechanistic view of cell-fate transitions governed by gain/loss of regulatory influence. Dynamic enrichment of regulators such as Klf4, Runx2, and Prrx1/2 was consistent with known trajectories and fate bifurcations in reprogramming.

Performance, Validation, and Numerical Highlights

Application to the mouse reprogramming dataset (14,953 genes, 576 TFs, eight clusters, 3,460 cells) was computationally tractable on standard workstations (<1GB RAM, ~1 hour per scMTNI run) and recapitulated canonical regulators, supporting biological validity. scMTNI, in both k-means and LDA analyses, robustly detected key pluripotency drivers (e.g. Esrrb, Klf2, Klf4), and transient regulators specific to intermediate or alternative fates. A matching score of 0.85 was reported for the core pluripotency cluster, providing quantitative support to the accuracy of edge recovery versus cell state identity.

Implications and Future Developments

scMTNI bridges important methodological and biological gaps by integrating regulatory priors with cell lineage information to model GRN dynamics at single-cell resolution, capturing regulatory inheritance and divergence. The approach offers immediate impact for studies in developmental biology, trans-differentiation, and disease progression, where lineage-specific regulatory rewiring is central. The modular implementation is compatible with unpaired multi-omic datasets and can be extended to integrate further modalities, or adapted for more complex lineage trees (e.g., cancer evolution, immune cell differentiation).

Future developments may include scaling to whole-organism single-cell atlases, accommodating spatial transcriptomics as an additional constraint, and further refinement of priors via machine learning over larger TF binding datasets. Extending the multi-task paradigm to model perturbation effects (e.g., CRISPR screens) is tractable within the current framework.

Conclusion

scMTNI presents a technically rigorous and flexible algorithm for dynamic, cell-type-resolved GRN inference from single-cell multi-omics data. Its explicit integration of lineage topology, motif and accessibility priors, and robust consensus strategies enable detailed exploration of regulatory dynamics during cell fate transitions. Applications to reprogramming systems confirm the method's ability to recover both canonical and novel regulators, providing a strong foundation for future mechanistic investigations in cell identity, development, and disease.