DynaMorph: self-supervised learning of morphodynamic states of live cells

Zhenqin Wu, Bryant B. Chhun, Galina Popova, Syuan-Ming Guo, Chang N. Kim, Li-Hao Yeh, Tomasz J. Nowakowski, James Zou, Shalin B. Mehta

TL;DR

This paper introduces DynaMorph, a novel self-supervised deep learning framework that analyzes live cell imaging data to automatically identify and classify distinct morphodynamic states. By combining label-free imaging with advanced computational techniques, DynaMorph enables quantitative analysis of cell behavior, even for challenging cell types like primary human microglia, paving the way for new discoveries in cell biology and disease research.

What Problem Does This Paper Address?

Understanding cell identity and function is deeply tied to their shape and movement (morphodynamics). However, automatically analyzing these dynamic states, especially in primary human cells where genetic labeling is often impractical, has been a significant challenge. Traditional methods rely on manual annotation, which is time-consuming and not scalable, or require specific molecular markers that can perturb cell behavior. This paper addresses the need for a robust, automated, and label-free method to quantify and classify complex cell dynamics, enabling high-throughput analysis of functional cell states.

Methods

DynaMorph employs a multi-stage approach:

  1. Label-Free Imaging: The system uses quantitative phase and retardance imaging (QLIPP) to capture cell density and anisotropy without the need for fluorescent labels. This provides rich, label-free information about cell morphology.
  2. Semantic and Instance Segmentation: A U-Net classifier, trained with sparse human annotations and augmented background data, is used for initial cell segmentation. This is followed by DBSCAN clustering to isolate individual cells and extract their masks.
  3. Self-Supervised Shape Encoding (VQ-VAE): A Vector Quantized Variational Autoencoder (VQ-VAE) is trained in a self-supervised manner to learn a compressed, interpretable latent representation of cell shapes. Crucially, this model is regularized by enforcing temporal continuity, ensuring that representations of consecutive frames are similar.
  4. Trajectory Feature Vector (TFV) Generation: For each tracked cell, the principal components (PCs) of its latent shape representation are concatenated with its instantaneous speed to create a Trajectory Feature Vector (TFV).
  5. Morphodynamic State Identification (GMM): A Gaussian Mixture Model (GMM) is used to cluster these TFVs, identifying distinct morphodynamic states within the cell population. This approach allows for the discovery of continuous spectra of states rather than discrete clusters.

Datasets

The study primarily uses primary human microglia isolated from brain tissue.

  • Training Data: Purified microglia without perturbation were used to train the DynaMorph model.
  • Test Data: Microglia treated with pro-inflammatory (IL17) and anti-inflammatory (IFN beta) cytokines, as well as glioblastoma (GBM) supernatant, were used to test the model and discover cell states.
  • Extension Analysis: Neural progenitor cells were also analyzed to demonstrate the generalizability of the framework.

Key Insights and Ideas

  • Label-Free Power: The success of DynaMorph hinges on its ability to extract rich morphological and dynamic information from label-free phase and retardance imaging, bypassing the need for genetic manipulation.
  • Self-Supervised Learning for Shape: VQ-VAE, regularized by temporal continuity, proves highly effective at learning generalizable and interpretable shape representations from complex cell images.
  • Shape Space as a Descriptor: The learned latent shape space captures fundamental cell properties like size and density, and importantly, its structure reflects the continuity of cell dynamics.
  • Beyond Static States: By analyzing trajectories and using GMM, DynaMorph can identify not just static morphodynamic states but also rare transition events between them, offering a dynamic view of cell behavior.
  • Microglia Complexity: The study reveals that microglia exhibit complex morphodynamic responses to disease-relevant stimuli like GBM supernatant, suggesting a heterogeneous population of states rather than a single response.

Key Results

  • DynaMorph successfully annotated morphodynamic states of human microglia using label-free imaging.
  • The VQ-VAE model achieved a significant data compression of over 680 times while maintaining high reconstruction fidelity.
  • The learned shape space, particularly the top PCs, showed strong correlations with interpretable geometric properties like cell size and density.
  • Analysis of microglia under different perturbations revealed distinct morphodynamic states, with IFN beta treatment leading to a state characterized by reduced density and migration speed, while GBM supernatant induced a broader distribution of states.
  • A two-state model (State 1: large, low-density, slow-moving; State 2: higher density, fast-moving) was identified, with control and IFN beta conditions showing enrichment in State 2 and State 1, respectively.
  • The framework demonstrated generalizability by successfully analyzing neural progenitor cells, showing distinct morphodynamic characteristics compared to microglia.

Limitations and Future Work

  • Segmentation Accuracy in Dense Cultures: The current segmentation and tracking accuracy is limited in very dense cell cultures, suggesting a need for methods that better handle overlapping cells.
  • Direct Correlation with Transcriptomics: While population-level correlations between morphodynamics and gene expression were observed, direct single-cell correlation remains an area for future work, potentially through paired measurements.
  • Biological Interpretation of States: The biological meaning of the identified morphodynamic states can be confounded by population heterogeneity and the abstract nature of the learned representation. Future work aims to develop a more quantitative vocabulary for describing cell dynamics.
  • Extension to Other Modalities: The authors suggest that DynaMorph can be extended to incorporate other quantitative measurements (e.g., metabolic states, cell cycle, protein localization) and applied to diverse cellular functions like differentiation, viral infection response, and apoptosis.

Summary updated

Abstract

The cell’s shape and motion represent fundamental aspects of the cell identity, and can be highly predictive of the function and pathology. However, automated analysis of the morphodynamic states remains challenging for most cell types, especially primary human cells where genetic labeling may not be feasible. To enable automated and quantitative analysis of morphodynamic states, we developed DynaMorph – a computational framework that combines quantitative live cell imaging with self-supervised learning. To demonstrate the fidelity and robustness of this approach, we used DynaMorph to annotate morphodynamic states observed with label-free measurements of density and anisotropy of live microglia isolated from human brain tissue. These cells show complex behavior and have varied responses to disease-relevant stimuli. DynaMorph generates quantitative morphodynamic representations that can be used to evaluate the effects of disease-relevant perturbations. Using DynaMorph, we identify distinct morphodynamic states of microglia polarization and detect rare transition events between states. The methodologies presented here can facilitate automated discovery of functional states of diverse cellular systems.
BibTeX
@misc{wu2020dynamorph,
  author = {Zhenqin Wu and Bryant B. Chhun and Galina Popova and Syuan-Ming Guo and Chang N. Kim and Li-Hao Yeh and Tomasz J. Nowakowski and James Zou and Shalin B. Mehta},
  title = {{DynaMorph: self-supervised learning of morphodynamic states of live cells}},
  year = {2020},
  doi = {10.1101/2020.07.20.213074},
  howpublished = {bioRxiv},
}
Figure