Cells are highly plastic. They grow, move and acquire new functions or cell fates in response to their environment and to treatment. Capturing these changes and identifying the molecular switches that control cell plasticity are major challenges in biomedical research. Meeting these challenges is essential if we are to better understand how diseases develop and how they can be treated effectively.
Single-cell sequencing has transformed biomedical research. Its success rests partly on the ability to analyse large numbers of cells and compare each experiment against the same reference genome or transcriptome. The thousands of genes or transcripts in these references act as a shared vocabulary.
Image analysis has lacked an equivalent standard. Researchers often have to choose a new set of features for each experiment or focus on selected time points because analysing videos at scale is difficult. This makes it harder to compare findings between studies. To make full use of live-cell imaging, researchers need both a common vocabulary for describing images and a tool capable of reading it.
In two complementary studies published in Nature Communications, researchers led by Professor Xin Lu at the Ludwig Institute for Cancer Research, University of Oxford, and Dr Felix Zhou at Vanderbilt University set out to address this problem. They established a live-imaging vocabulary called the SAM phenome and developed a companion reading tool called SPOT. Applying SAM-SPOT to millions of 2D and 3D live-cell images, they found that it could outperform feature sets generated by machine-learning models.
A standardised vocabulary for cell behaviour over time
In the first study, the researchers defined 2,185 measurable features describing cell shape, appearance and motion. Together, these features make up the SAM phenome.
They also developed the SAM Phenotype Observation Tool, or SPOT, to measure these features and follow how they change over time. By representing every experiment using the same feature set, researchers can compare, combine and cluster imaging data without having to design a new analytical pipeline for each study.
In benchmarks using more than one million cell images, the SAM phenome performed better than several feature sets generated by deep-learning models. It also remained interpretable: each SAM feature describes a named, measurable property that researchers can examine directly.
Connecting organoid behaviour with gene activity
In the companion study, the researchers applied SAM-SPOT to organoids, three-dimensional cell cultures that reproduce selected structural and functional characteristics of organs.
The framework detected differences in organoid growth, shape and responses to drugs, genetic changes and culture conditions. Some of these changes were difficult to identify using conventional methods.
The researchers then combined the imaging measurements with single-cell RNA sequencing from the same organoids. This revealed links between observed behaviour and specific biological processes, including WNT signalling, connecting changes in tissue development with underlying gene activity.
SAM-SPOT works with standard bright-field images, so it does not require fluorescent labels or force researchers to decide in advance which proteins or behaviours to monitor. Fluorescent tagging remains valuable for studying particular molecules, but it limits the number of targets that can be observed at once and can sometimes affect cell behaviour. By measuring many features simultaneously without labels, SAM-SPOT can identify complex or unexpected changes while cells and tissues are followed over time.
The researchers suggest that the framework could support large-scale genetic and drug screens using live-cell imaging. It could also be adapted for digital pathology and spatial imaging, where a shared vocabulary may help researchers identify and compare patterns of tissue organisation across samples.