APPLYING IN SILICO LABELING VIA TRANSFER LEARNING TO DISSECT ORGANELLE - ORGANELLE SPATIAL DEPENDENCIES

How organelles act in concert to shape and enable cell function is a fundamental question in cell biology. Progress in the field is hindered by lack of systematic tools to dissect spatiotemporal interorganelle interactions, mainly due to technical limitations in simultaneous labeling of multiple organelles within the same living cell. We combine emerging computational techniques and a publicly available comprehensive dataset of 3D cell imaging to validate known relations and to raise new hypotheses regarding inter-organelle spatial dependencies.

Modern machine-learning techniques were recently shown to successfully extract the locations of organelles within the cell from label-free images, a technique called “in silico labeling”. Transfer learning is a machine learning technique where computational models trained for one task are used as a starting point to train a new model to a different but related task. The intuition for the success of transfer learning stems from the understanding that re-training models using better initial states can provide benefits when training data are limited in volume. We use this property as means to measure spatial dependencies between different organelles. If an in silico labeling model trained to localize an organelle is improved by transferring a model trained to localize a different organelle, than we say that the information encoded in the mapping to the latter organelle is useful for the prediction of the former organelle. Such comparison between two models defines an asymmetric link that encodes a spatial dependency between the latter and the former organelles. We applied this methodology to construct an inter-organelle spatial dependencies network by integrating the pairwise relations between 13 organelles in 3D imaged endogenously labeled human-induced pluripotent stem cells available by the Allen Institute of Cell Science.

Our preliminary results validate several known relations between organelles and propose new predictions that should be validated experimentally. For example, both the nuclear envelope and the endoplasmic reticulum (ER) contribute to the prediction of the golgi apparatus, and the nuclear envelope contributes to the prediction of the ER. The golgi and the tight junctions are linked bi-directionally. The plasma membrane contributes to the prediction of adherens junctions. No organelle enhances the prediction of desmosomes.

Ultimately, our project will provide a comprehensive and validated methodology for predicting organelle-organelle interactions, bypassing long-standing technical barriers in microscopy. The establishment of a rich resource of predicted organelle-organelle interaction networks can providing a major leap towards the “holy grail” of cell biology – inclusive understanding of cells as integrated complex systems.