Generalizable and explainable deep learning for morphological profiling of cells : methods and applications

Abstract

The field of cell biology has been revolutionized by advanced microscopy, which has significantly improved image quality and speed. Researchers can now explore the intricate structural and functional aspects of single-cell morphology, transforming single-cell imaging technology into data-driven science, and thus enabling comprehensive analysis of cell health, disease mechanisms, and responses to chemical and genetic perturbations. With the growth of open image repositories and advancements in machine learning, researchers can extract morphological fingerprints or profiles and enable in-depth analysis of cellular morphology. The process of extracting morphological information in high dimensions and utilizing it to interpret the data can however be a daunting task. It traditionally requires meticulous fine-tuning of feature selection and careful consideration of statistical data analysis techniques, considering their advantages and limitations. This thesis aims to address the challenge of manual feature extraction and high-dimensional analysis by introducing a novel integrative unsupervised deep-learning framework called MorphoGenie, focusing on three crucial tasks: The first task involves adopting a deep learning-based unsupervised disentangled learning approach for single-cell morphological profiling and downstream analysis. This approach aims to capture biological information inherent in the dataset. Secondly, disentangled representations encode interpretable information about single-cell image features. A novel strategy for interpreting the latent representation is proposed. The third task focuses on employing the disentangled representations for generalization across different unseen imaging modalities. By leveraging the morphological insights, the model learned during training, it becomes possible to apply the knowledge to new imaging modalities effectively. The thesis introduces a metric for assessing the disentanglement of single-cell morphological features. This serves as a quantitative measure to evaluate the effectiveness and interpretability of the disentangled representations. The performance is evaluated by comparing it with state-of-the-art autoencoders in various aspects, including downstream analysis and reconstructions. This evaluation is conducted on a diverse range of datasets, encompassing different imaging modalities, imaging conditions (adherent and suspension cells), varied shape morphologies, and biological conditions (discrete cell types and continuous cellular progression). The best-performance models assessed by disentanglement metric are chosen from each dataset and are then used for testing generalizations. The integrative unsupervised deep learning-based framework is a valuable tool for large-scale image morphological profiling, revealing insights into phenotypical and morphological variations across diverse conditions. It eliminates the need for retraining, enabling efficient study of cellular morphology and behaviours in diverse biological contexts. As part of the thesis research, a cutting-edge laser-scanning imaging cytometry platform was used to capture real-time, high-throughput images of over 10,000 cells per second. This platform allowed detailed analysis of individual cells in terms of their physical properties, enabling a comprehensive study of viral infection dynamics. By using multi-contrast bright-field and quantitative phase imaging (QPI), it was possible to obtain detailed morphological information without the need for labelling. Results showed that this approach was effective in assessing infections in different cell types at various stages, highlighting its potential for use in diagnostics and therapeutic studies. Furthermore, the dataset derived from this study was employed as an additional source of validation for the MorphoGenie framework.