Morphodynamical cell state description via live-cell imaging trajectory embedding

Abstract

Time-lapse imaging provides powerful insight into the dynamical response of cells to perturbation, but the quantitative analysis of morphological changes over time is a challenge. Here, we exploit the concept of "morphodynamical trajectories" to analyze cellular behavior using morphological feature trajectory histories, rather than the common practice of examining morphological feature time courses in the space of single-timepoint (snapshot) morphological features. Our morphodynamical trajectory embedding analysis yielded quantitative and descriptive models of future time points based on the extended history information of MCF10A mammary epithelial cells treated with a panel of ligands. The trajectory analysis constructs a shared morphodynamical cell-state landscape, where the response of MCF10A cells induced by various extracellular signals is characterized by ligand-specific regulation of state transitions. Additionally, we show that including trajectories in single-cell morphological analysis enables (i) systematic characterization of cell state trajectories, and (ii) better separation of phenotypes and more descriptive models of ligand-induced differences as compared to snapshot-based analysis. This morphodynamical trajectory embedding is broadly applicable for the quantitative analysis of cell responses via live-cell imaging across many biological and biomedical applications.

Machine learning inference of continuous single-cell state transitions during myoblast differentiation and fusion

“…Specifically, using machine learning for quantitative characterization requires extracting features that can capture the state transition and avoiding features that may confound the quantitative characterization of the process (e.g., avoiding local cell density in characterizing the differentiation process). In our case, and in agreement with other studies (Copperman et al, 2021; Wang et al, 2020; Wu et al, 2022), integration of multiple dynamic features encoding the temporal changes were necessary to continuously measure a biological process.…”

“…We combined live cell imaging and machine learning to infer the differentiation state of single cells during the process of muscle precursor cell differentiation. Many studies highlight the rich information encapsulated in single-cell dynamics that, with the aid of supervised or unsupervised machine learning, enable effective identification of sub-populations and discrimination of perturbations (Choi et al, 2021; Goglia et al, 2020; Jacques et al, 2021; Jena et al, 2022; Kimmel et al, 2018; Valls & Esposito, 2022), that cannot be inferred from static snapshot images (Copperman et al, 2021; Wang et al, 2020; Wu et al, 2022). For example, approaches that rely on static snapshots make it extremely hard to infer trajectories that deviate from the mainstream cell state progression because they are confounded by cell-to-cell variability.…”

“…For example, whether the difference in scores between 0.3 and 0.4 has the same phenotypic magnitude as between scores of 0.6 and 0.7. This limitation is also common to approaches that use non-linear dimensionality reduction (Copperman et al, 2021; Eulenberg et al, 2017; Jacques et al, 2021; Rappez et al, 2020; Stallaert et al, 2022; Wang et al, 2022; Wu et al, 2022) and could also limit unsupervised state representations that can be dominated by features that do not relate to the cell state (Copperman et al, 2021; Jacques et al, 2021; Wang et al, 2022; Wu et al, 2022). Still, the monotonicity property holds, e.g., a differentiation score of 0.4 is predicted to be more advanced along the differentiation trajectory than a differentiation score of 0.3.…”

“…Live cell imaging offers a solution to this challenge by enabling dynamic monitoring and extraction of temporal information at the single-cell resolution. However, unsupervised approaches applied to live cell imaging may still be confounded by variability factors that are unrelated to the state transition of interest (Copperman et al, 2021;Wang et al, 2022).…”

“…However, the capacity to identify single cell trajectories that deviate from the most common progression, is limited in this approach, due to heterogeneity (Schroeder, 2011). Live-cell imaging offers a solution to this challenge by enabling dynamic monitoring and extraction of temporal information at the single-cell resolution, with the caveat that unsupervised modeling may consider extrinsic factors that are unrelated to the state transition, which may confound proper modeling of the continuous process (Copperman et al, 2021; Wang et al, 2022; Wu et al 2022).…”