CellPhe 2: Machine Learning-based Phenotyping To Predict Cell Fate In Heterogeneous Populations

Within the BBSRC remit "understanding the rules of life", advancements in high-throughput time-lapse imaging has led to a proliferation of large datasets reporting temporal changes in cell behaviours in mechanistic functional studies, target validation, and phenotypic screens. These large and complex datasets can answer fundamental questions in biology but require more advanced tools for analysis.

As part of a BBSRC-funded NPIF PhD studentship in AI, we developed CellPhe, a software package that enables phenotyping of cells in time-lapse microscopy datasets. Whilst some other software systems allow visualisation of the time series for a limited number of size and shape metrics, CellPhe is unique in that the time series for over 70 features, including movement, texture and density, are calculated for each tracked cell. Quantitative attributes are extracted from each of these time series and used in machine learning algorithms to classify cell subpopulations, thereby filling a fundamental gap in the bioimaging field. In this project, we aim to significantly enhance CellPhe's capability whilst making the software more widely accessible to the bioimaging community.

To harness CellPhe's capability for more complex datasets of heterogeneous cell populations, we will employ a novel analytical approach where we consider "tracklets" of imaging data rather than the complete time series. These short segments of the time series will allow interesting short-term behaviour to be identified that would be missed by characterising the time series in full. This approach will be used to identify key cellular processes that may induce phenotypic shifts in response to treatment, e.g. timing of cell division, number and type of transient cell-cell interactions, migration, local density, etc.

A common limitation with time-lapse imaging data is the inability to accurately segment and track individual cells across multiple frames, particularly in more confluent populations. To address this issue when it arises in specific datasets, we will explore the alternate utility of analysing a population of cells across a whole field of view without segmentation, using machine learning to separate behaviours based on whole image metrics.

Furthermore, to demonstrate utility for discrimination of responding/non-responding populations following exposure to specific interventions (e.g. during drug treatment or CRISPR screening), we will also run the data in reverse to determine the features that provide early indication of particular cell behaviours. This approach will aid classification models for the prediction of cell fate, e.g. early detection of differentiation, or drug response, and could allow analysis of phenotypic imaging data to be combined with additional downstream assays, e.g. spatial transcriptomics, to relate imaging data to function.

We will achieve our goal of making CellPhe more accessible by integrating it with the widely used state of the art open-source image analysis package, CellProfiler Analyst. Its companion program CellProfiler, led by our international partner, is recognised as the leading solution for image segmentation and cell tracking. The collaboration will allow segmentation and tracking to be integrated with automated analysis of the tracked cells. Further collaborations with two large pharmaceutical companies will provide access to industrially relevant high throughput imaging datasets for testing these new analytical capabilities. Our proposal thus addresses the BBSRC theme "AI to transform bioscience research and discovery", opening up a significant new opportunity for analysis of high dimensional time series datasets applicable to cellular bioimaging and beyond.