DIISCO: A Bayesian framework for inferring dynamic intercellular interactions from time-series single-cell data

Cameron Y. Park, Shouvik Mani, Nicolas Beltran-Velez, Katie Maurer, Satyen H. Gohil, Shuqiang Li, Teddy Huang, David A. Knowles, Catherine J. Wu, Elham AzizibioRxiv 2023 · DOI

Abstract

Characterizing cell-cell communication and tracking its variability over time is essential for understanding the coordination of biological processes mediating normal development, progression of disease, or responses to perturbations such as therapies. Existing tools lack the ability to capture time-dependent intercellular interactions, such as those influenced by therapy, and primarily rely on existing databases compiled from limited contexts. We present DIISCO, a Bayesian framework for characterizing the temporal dynamics of cellular interactions using single-cell RNA-sequencing data from multiple time points. Our method uses structured Gaussian process regression to unveil time-resolved interactions among diverse cell types according to their co-evolution and incorporates prior knowledge of receptor-ligand complexes. We show the interpretability of DIISCO in simulated data and new data collected from CAR-T cells co-cultured with lymphoma cells, demonstrating its potential to uncover dynamic cell-cell crosstalk.

Introduction

Cells do not function in isolation; they coordinate complex biological processes through continuous chemical signaling mediated by ligands and receptors. While single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to observe gene activity, inferring the dynamic "conversations" between cells remains a significant computational hurdle. Most existing tools provide only static snapshots of communication and are heavily dependent on fixed databases, which limits their ability to identify novel interactions or track how the strength and direction (activating vs. inhibitory) of crosstalk evolve during treatment or disease progression.

The DIISCO Framework

To address these limitations, we developed DIISCO (Dynamic Intercellular Interactions in Single Cell transcriptOmics), a generative Bayesian framework built on Gaussian Process Regression Networks (GPRNs). DIISCO treats cell type proportions over time as a multi-output regression problem, assuming that the frequency of a specific cell type at time $t$ is influenced by other cell types through a time-varying interaction matrix $W(t)$.

Overview of the DIISCO framework, illustrating the integration of temporal cell proportions and prior knowledge to infer dynamic interactions.

By utilizing Gaussian Processes (GPs), DIISCO can model complex, non-linear biological changes without requiring a predefined functional shape. The framework incorporates prior biological knowledge from databases like OmniPath to focus on plausible interactions while allowing the model to learn the actual strength and sign of these interactions directly from the longitudinal data. Variational inference is employed to estimate posterior distributions, providing a robust measure of uncertainty for all predictions.

Experimental Validation and Results

DIISCO was validated using both synthetic simulations and a novel experimental dataset. The simulation involved five cell types with seven distinct interaction dynamics (e.g., transient, increasing, decreasing), where DIISCO successfully recovered the "ground truth" interaction curves.

For biological validation, we co-cultured CAR-T cells with MEC1 (lymphoma) cells and performed scRNA-seq at ten time points over a 24-hour period. This longitudinal study allowed us to observe the transition of T cells from activation to exhaustion as they interacted with cancer cells.

Analysis of CAR-T and MEC1 co-culture data, showing the experimental workflow, cell clustering via UMAP, and marker gene expression profiles.

Key findings from the CAR-T data included the identification of the ICAM1-IL2RA interaction as a top-ranked pair, which is consistent with its known role in inflammatory responses in B-cell chronic lymphocytic leukemia. Furthermore, DIISCO demonstrated that interaction strengths between exhausted T cells and cancer cells increased over time, mirroring the biological reality of the mounting cellular response. Benchmarking against standard tools like CellChat and CellPhoneDB showed that DIISCO’s temporal predictions correlated significantly better with actual gene expression data over time.

Limitations and Future Directions

The precision of DIISCO is inherently linked to sampling density; sparse time points or large temporal gaps can lead to increased model uncertainty. Additionally, because the model relies on shifts in cell type proportions, it may struggle to detect intense interactions in stable tissues where cell numbers do not fluctuate significantly. Future work will focus on integrating spatial transcriptomics data, allowing DIISCO to account for physical proximity in addition to temporal co-evolution, further refining our understanding of the cellular "conversation."

Summary updated

Abstract

Characterizing cell-cell communication and tracking its variability over time is essential for understanding the coordination of biological processes mediating normal development, progression of disease, or responses to perturbations such as therapies. Existing tools lack the ability to capture time-dependent intercellular interactions, such as those influenced by therapy, and primarily rely on existing databases compiled from limited contexts. We present DIISCO, a Bayesian framework for characterizing the temporal dynamics of cellular interactions using single-cell RNA-sequencing data from multiple time points. Our method uses structured Gaussian process regression to unveil time-resolved interactions among diverse cell types according to their co-evolution and incorporates prior knowledge of receptor-ligand complexes. We show the interpretability of DIISCO in simulated data and new data collected from CAR-T cells co-cultured with lymphoma cells, demonstrating its potential to uncover dynamic cell-cell crosstalk.
BibTeX
@misc{park2023diisco,
  author = {Cameron Y. Park and Shouvik Mani and Nicolas Beltran-Velez and Katie Maurer and Satyen H. Gohil and Shuqiang Li and Teddy Huang and David A. Knowles and Catherine J. Wu and Elham Azizi},
  title = {{DIISCO: A Bayesian framework for inferring dynamic intercellular interactions from time-series single-cell data}},
  year = {2023},
  doi = {10.1101/2023.11.14.566956},
  howpublished = {bioRxiv},
}
Figure