0
$\begingroup$

Usually, cell types are found by clustering cells in expression space (or some lower dim. projection of it), and calculating differentially expressed (DE) genes. These DE genes are then used to annotate the cluster to a particular type. Cell states are then thought of as a 'fine-grained' view on this cluster, making some reference to a biological process.

However, this means that a particular cell state can never be present across cell types, whereas we do think of them like this. For example, "being at a certain point along the cell cycle" can be thought of as a state, or "being under stress", and cells of different types can be in these states simultaneously. Is there some principled way to extract such states from single cell gene-expression data? It should be able to annotate different, far removed regions in expression space as being in the same state, so probably cannot be based on distances in expression space.

$\endgroup$
3
  • 1
    $\begingroup$ It kind of sounds to me like you're thinking only like a computer scientist and not like a biologist. If you want to know where a cell is in the fell cycle, consider looking at cell-cycle related genes. If you want to look at stress, look for expression of stress-associated genes. $\endgroup$
    – Bryan Krause
    Jun 8, 2022 at 17:33
  • $\begingroup$ That is true, I am approaching this as more of a computational problem. That said, I think there would be added value in being able to detect novel states, not just ones for which you know a set of marker genes. For example, there might be a state defined by three genes (A, B, C) being expressed together. It could be the case that these three are only expressed together in distinct but disjoint locations in expression space. I was wondering if there are known methods/algorithms for detecting such states. $\endgroup$
    – Abelaer
    Jun 9, 2022 at 8:44
  • $\begingroup$ I would expect that the kind of correlation structure that you describe would be amenable to multivariate analysis. As a very simple-minded approach, you would expect that PCA would find a projection in which (A, B, C) are all strongly loaded and other genes are at the noise level if their signal is strong enough and your data are properly normalized (how to normalize is a separate question), even in the presence of other large expression signals like cell cycle etc. Can you add more specifics about what exactly you have tried? $\endgroup$ Jun 9, 2022 at 18:11

1 Answer 1

1
$\begingroup$

My suggestion is that you rethink the premise of your question. Projections and distance comparisons can be used to define cell states that apply to multiple cell types.

Here is an example. A and B are two cell types. S and T are two cell states. x and y are two genes used to define these states. The plot below shows expression of gene x on the x-axis and gene y on the y-axis. The vertical line separates cell types. The horizontal line separates cell states.

                gene y
                ^
                |    
    state: S    |    state: S    
    type:  A    |    type : B    
                |    
<==============================> gene x
                |    
    state: T    |    state: T    
    type:  A    |    type : B    
                |    
$\endgroup$
1
  • $\begingroup$ Thanks! This is a very useful diagram, and indeed how I had been thinking about it. I've accepted your answer as I think it illustrates the point nicely. $\endgroup$
    – Abelaer
    Jun 28, 2022 at 8:08

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .