In recent years, people have stopped being satisfied with bulk measurements that tell us only what an average cell might look like. Single-cell sequencing made it possible to look at cells separately and to see that a developing or diseased tissue is actually a mosaic. Workhorse methods such as scRNA-seq, scATAC-seq and the newer spatial transcriptomic approaches let us line up three kinds of information in the same biological system: which genes are active, which chromatin regions can be accessed, and where a given cell actually sits in the tissue. When these tools are used on embryos, investigators can follow the opening of chromatin through organ formation, point to transcription factors that push cells toward one fate rather than another, and even notice metabolic shifts that occur while immune cells are maturing. Disease studies tell a similar story, but from the pathological side: single-cell data make it easier to say “this subgroup of cells is where the problem starts,” and to describe the metabolic or damage-related patterns those cells create. None of this is effortless—batches do not always match, per-sample costs are still high, and downstream analysis is not trivial. This is why people are trying to combine different single-cell readouts, to let machine-learning models handle some of the complexity, for example through multi-omics integration that links transcriptomic and chromatin features to spatial context, and to bring down the overall price, so that the same logic can be used in real personalized medicine rather than only in research papers.
邹大有 (Wed,) studied this question.