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.
Availability DIISCO is publicly accessible at https://github.com/azizilab/DIISCO_public. All data will be deposited to GEO upon publication.
Competing Interest Statement
C.J.W. is an equity holder of BioNTech and receives research funding from Pharmacyclics.
Footnotes
↵{cyp2111{at}columbia.edu,ea2690{at}columbia.edu}
13 We also allow for a bias term b(t) such that Y (t) = W (t)f(t) + b(t) + ϵ(t) where b(t) is centered at 0 and has a kernel like that of W (t) but without any interaction regularization. For brevity, we omit it as it is equivalent to extending the model dimension by one, ignoring the last coordinate of y(t), and setting the last coordinate of f(t) to be a GP centered at 1 with infinite length-scale.
14 The arrows 𝒲o→ 𝒲u and ℱo → ℱu represent tractable distributions that can be computed analytically and can be sampled due to the properties of Gaussian processes. Thus, the model only requires tractable operations that lead to the joint distribution described.
15 In the case when we are dealing with proportions and not raw counts this also ensures that we avoid a trivial solution due to theΣk yk(t) = 1.
