A framework to integrate live-cell imaging with single-cell sequencing and learn how cells adapt to new environments

点击翻译按钮获取中文摘要

Abstract Exposing cancer cells to a new environment typically influences their growth. For some cells, a moderate growth inhibition is followed by adaptation and return to normal growth rates. Others initially experience a near-complete cytostatic phenotype, only to explode in their growth during later generations, reaching growth rates well beyond baseline. This implies that one can reach opposite conclusions about the relative fitness of two cell lineages, solely depending on timing of measurement. This has implications for the time window of therapeutic success, in light of the fact that virtually all pre-clinical drug screening studies measure growth rate at a single, well defined timepoint – typically 72 hours after exposure. Despite this shortcoming, measurements of cell fitness at multiple timepoints are impractical for large-scale studies. Our long-term goal is the development of a new class of temporal biomarkers that extrapolate from a cell's transcriptome how fit its descendants will be over multiple generations. As a next step towards this goal we propose a feasibility study to collect training data of su!cient temporal reach and cellular resolution to evaluate the predictability of cell fitness. With a broad record of integrating various omics- and imaging platforms and as the developers of widely deployed drug response metrics, our team brings complementary expertise to integrate live-cell imaging with single cell sequencing for deep learning. We will record how cells divide, migrate and die, linking the recorded phenotypic di"erences between cells to di"erences between their transcriptomes. Aim 1 will use live-cell imaging to characterize the cell cycle of cancer cell clones as they adapt to new environments. Hereby we define a growth condition as the combination between founding cell and micro-environment. We hypothesize that time emphasizes di"erences between growth conditions, i.e. that cell cycle progression profiles from distinct growth conditions diverge as their cells converge on a specific path of adaptation. This temporal evolution of cell adaptation will inform which generation is optimal for single-cell RNA sequencing, namely when cell counts are su!ciently high, but the path of adaptation is not yet phenotypically evident. In aim 2 we will test the potential of the transcriptome to predict this path. To achieve this we will use a three-layered approach to map sequenced- and imaged cells in-silico. Hereby biological variability – emerging from multiple growth conditions – acts as an additional barcode during sequencing. This linking will not only match the sequenced cell's transcriptome to the one phenotype of the corresponding imaged cell, but also to adaptive phenotypes of all its ancestors. The outcome of these two aims will be training data to learn: (i) whether a snapshot of a transcriptome has the potential to forecast speed and success rate of cell cycle progression; (ii) the temporal limitations of such predictions and (iii) how much more data will be needed to train a deep neural network to make such forecasts. Integration with live-cell imaging opens the door to truly leverage the suitability of single cell sequencing for deep learning in a new way – not for solving technical challenges like segmentation and tracking, but for interpretation of genomic information. We aim for our e"orts to reduce the one-order-of-magnitude temporal chasm between in-vitro cancer cell adaptations and the time required for clinically relevant phenotypes to emerge. 2