(648d) A Deep Learning Approach for Predicting Cell Motility States and Evaluating Cellular Heterogeneity | AIChE

(648d) A Deep Learning Approach for Predicting Cell Motility States and Evaluating Cellular Heterogeneity


Maity, D. - Presenter, Johns Hopkins University
Jayatilaka, H., Johns Hopkins University
Zamponi, N., Weill Cornell Medicine
Giri, A., Johns Hopkins Physical Sciences - Oncology Center
Agrawal, A., Johns Hopkins University
Walston, J., Weill Cornell Medicine
Wirtz, D., Johns Hopkins University
Phillip, J., Johns Hopkins University
Populations of cells typically display dynamic and heterogeneous phenotypes. In this study, we use ~24,000 single cell trajectories and over 400 different biological conditions, to characterize patterns of cell motility and provide an effective approach to quantifying cellular heterogeneity.

First, to unify cell trajectories, we developed a custom inverse Monte-Carlo technique that allows the quantification of cell motility across timescales, based on total durations and time intervals for acquisition. After testing, we identified a total duration of 8 hours and an acquisition rate of 5 minutes as an appropriate time-scale. We performed rigorous validation and sensitivity analysis to compare higher-order parameters such as mean squared displacements (MSD) and diffusivities, to check the concordance among trajectories. Using machine learning approaches, such as unsupervised hierarchical clustering and k-means clustering we demonstrated that cells can be classified into various motility states based on spatio-temporal patterns. Upon clustering, we robustly identified 25 clusters. We then validated the stability of our clustering using t-stochastic neighbor embedding (t-SNE) and calculating the inertia and silhouette values. In addition, our analysis shows that the 25 clusters can be further grouped based on similarities, which was also observed in t-SNE plots. We investigated the differences between clusters via a comparison of higher order parameters, such as MSD, persistence (temporal), diffusivity and anisotropy (spatial persistence). Analysis was performed across multiple cell types, cancerous/non-cancerous, mesenchymal/non-mesenchymal, epithelial/fibroblast-like, and cells migrating through various geometries of 2D/3D. We also investigated cases where cells have been treated with various pharmacological perturbations which include cytoskeletal modulators i.e., actin /microtubule and myosin-activity modulators. Based on these analyses, we demonstrate that cell motility heterogeneity is linked to the fractional distribution of cells among motility states as a function of conditions/perturbations, which can be used to classify cells and compute the extent of cellular heterogeneity based on the entropy of the distributions.

Lastly, in order to build on our approach to include new cells, we developed methods that can be used to predict the state of cell motility regardless of acquisition rate and total acquisition times. To predict the state of motility with acquisition rates of 5 minutes, we trained a machine learning algorithm, namely, Support Vector Machine, which showed a validation accuracy of ~97%. Furthermore, to predict the motility state for cells having different acquisition rates (not equal to 5 minutes), we trained a long-short term memory (LSTM) network, demonstrating an average validation accuracy of ~95% across 25 Clusters.

Conclusion: By analyzing >24,000 single cell trajectories and across multiple different biological conditions, we demonstrate that cell motility patterns can define multiple states. To enable classification of new cell trajectories into these states we devised two ML-based approaches. Our prediction platform provides a way to define cell motility states and enables cell classification without the need to re-cluster cell trajectories. Our system also provides a way to assess the efficacy of various perturbations and hence can help improve drug discovery and development as a screening platform.