Causality-aware foundational topic modeling for single-cell multi-omics and disease mechanisms

The National Institute of General Medical Sciences awarded Emory University $423,357 on August 1, 2026, under the Biomedical Research and Research Training program (CFDA 93.859) to develop a scalable, causality-aware foundational topic modeling framework for single-cell multi-omics data.

The project addresses critical limitations in existing representation learning methods for high-dimensional single-cell multi-omics datasets by developing interpretable topic representations enriched in causal relational semantics. Specific deliverables include a multimodal topic model that encodes heterogeneity and directionality of relationships among cells and multi-omic features, a scalable continual learning algorithm to reconcile topic representations from atlas-scale datasets into a unified foundational topic space, and a causal inference framework to identify putative causal topics and biological programs underlying cell state transitions. These representations will support diverse analytical tasks, knowledge mining across massive datasets, and identification of causal biological programs underlying developmental and pathological state transitions.

Work is performed at Emory University in Atlanta, Georgia, with a period of performance from August 1, 2026, through March 31, 2031. The award is classified as a Project Grant.

Abstract

Single-cell (SC) multi-omics technologies enable the study of cellular states at unprecedented resolution across multiple molecular layers, offering unique opportunities to uncover molecular mechanisms driving biological development and disease progression. To analyze these high-dimensional, sparse, noisy, and heterogeneous data, representation learning methods have been developed to generate lower-dimensional, semantically enriched embeddings applicable to various downstream tasks. However, existing methods face critical limitations: they often produce embeddings that lack interpretability, do not scale to massive datasets, fail to capture heterogeneous and directional relationships among cell states and omic features, and remain restricted to associative rather than causal discovery. To address these gaps, i propose to develop a scalable, causality-aware foundational topic modeling framework for SC multi-omics data. This framework aims to generate versatile and interpretable topic representations enriched in various causal relational semantics at both cell and omic feature levels. These representations will support diverse analytical tasks, knowledge mining across massive datasets, and identification of causal biological programs underlying developmental and pathological state transitions. Specifically, i will: (1) develop a multimodal topic model that leverages heterogeneous directed graphs to encode the heterogeneity and directionality of relationships among cells and multi-omic features into topic representations; (2) develop a scalable continual learning algorithm to reconcile topic representations from atlas-scale datasets into a unified foundational topic space that can be incrementally updated and expanded; (3) design a causal inference framework to identify putative causal topics and their implied biological programs underlying cell state transitions, with statistical guarantees provided through power analysis and sample size estimation; and (4) collaborate with wet-lab partners to apply the framework to investigate molecular mechanisms driving cellular state transition during the development of diseases such as neurodegenerative disease and cancer. Putative causal biological programs will be opportunistically validated through perturbation experiments in organoids and ipsc-derived models. The successful completion of this project will establish the very first scalable, causality-aware foundational topic modeling framework for SC multi-omics, yielding high-quality, unified, and interpretable topic representations across datasets. These advances will enhance downstream analytical performance, enable transparent and human-interpretable analyses, and facilitate knowledge discovery from massive data resources. Finally, all algorithms will be disseminated as user-friendly, open-source software broadly to benefit the single-cell research community in facilitating mechanistic discovery, accelerating hypothesis generation, and promoting translational applications such as biomarker discovery and therapeutic target identification.