Not sisters, under the skin

Ever since we’ve been able to look at the internal components of cells, we’ve become more and more fascinated by the fact that individual cells are — well — individual.  Two genetically identical cells sitting next to each other in a dish may harbor very different sets of messenger RNAs and proteins (or other components), and may therefore respond differently to a stimulus.  We’ve seen this in the context of apoptosis in mammalian cells, and also bacterial responses to antibiotics.  Where do these differences come from?

The dominant story, so far, is that the variation in mRNA and protein levels we see is due to gene expression noise.  In a normal cell, you have at most 2 copies of any given gene, and each gene may be either in an “on” state for transcription (producing mRNA) or an “off” state.  This can produce bursts of mRNA production (think of the sudden stream of cars that comes through a traffic light when it turns green, then stops when the light turns red), which in turn can produce even larger bursts of protein production.  This will make protein levels fluctuate randomly over time, and different cells are not very likely to fluctuate in synchrony.  So there are bound to be differences between individual cells: noise is a fact of life.

But is gene expression noise the only source of fluctuations?  A recent paper (Huh and Paulsson, 2010, Non-genetic heterogeneity from stochastic partitioning at cell division, Nature Genetics doi:10.1038/ng.729) argues that cell-to-cell differences created at cell division when components are unequally split between the two daughters (called partitioning error) may be just as important as — or more important than — gene expression noise.

This conclusion may surprise people who’ve been following the noise field closely, because the impression recent papers have given is that gene expression noise can explain the fluctuations we see extremely well, with no need to invoke cell division as a source of noise.  This impression primarily comes from a particular type of experiment which shows that increasing the rate of gene expression reduces the normalized level of noise.  That seems reasonable, right?  If the noise comes from fluctuations in gene expression, then doubling gene expression should halve the relative variance.  Doubling the frequency of a traffic light turning green would make the flow of traffic more consistent, at least in one direction.

Well, yes.  The problem is that increasing gene expression should also reduce fluctuations that are due to partitioning error. Huh and Paulsson show that, unless something special is going on — a caveat we’ll return to in a moment —both sources of fluctuations depend in exactly the same way on the average total number of molecules.  If you increase gene expression, you have more of the component you’re following — whether it’s mRNA or protein — and the larger the number of molecules, the harder it is to have a dramatically non-equal distribution between the two daughters.  Not impossible — it’s not impossible to get 10 heads in a row by chance when tossing a coin, it’s just less likely than getting 2 heads in a row. What this means is that a model ignoring or idealizing partitioning error and focusing only on the stochasticity of synthesis fits the available data in many papers very well; but so does a model that does the exact opposite.  [Note to self, and to all modellers: just because a model fits the data doesn’t mean it’s right.]

“Unless something special is going on”. What does that mean?  Some types of cellular components have special mechanisms to ensure that they are evenly divided between daughter cells.  Chromosomal DNA is the most obvious example, but there are a lot more: plasmids, carboxysomes and the Golgi, for example.  Other components, even organelles, don’t. (I recently wrote about a situation where cell-to-cell variation was traced to unequal partitioning of mitochondria, though this may be unusual.) Huh and Paulsson show that in the absence of such special mechanisms the randomness introduced by partitioning and the other largely ignored source of variation, degradation, is in the range of 33-75% of the total variance.  That seems important enough not to ignore.

It’s also worth noting that any component that clusters in lumps will be more randomly distributed after cell division than you would expect given the number of individual components. And yet another key point about partitioning error is that it’s even harder to correct than gene expression noise.  You only have one cell cycle in which to attempt to get back to normal, otherwise the error will be propagated further, to your daughters. And since the errors can be large in either direction, you need to be able to control both synthesis and degradation to have a chance of fixing them.  So if you have partitioning errors, they’re likely to be important.

Are there better ways to distinguish gene expression noise from partitioning error, beyond manipulating gene expression?  Huh and Paulsson show, regretfully, that the experimental methods available are severely limited in this regard.  Measuring the full distribution of component numbers doesn’t really help, and even watching the distribution of GFP-labeled molecules during cell division isn’t useful: if the component you’re studying is abundant enough that you can track it using normal fluorescent imaging techniques, the partitioning error in that component will be minimal.  But its fluctuations could still be driven by partitioning error. For example, the mRNA molecules that produce the GFP-labeled protein may be present in very low copy numbers and segregated randomly at cell division; even in the absence of gene expression noise, this would lead to significant variation in the levels of the GFP-labeled protein.

There’s hope, though.  Although you can’t use experiments in which you increase synthesis rates to show that gene expression is the source of the fluctuations, you can use such experiments to identify the cellular component that is primarily driving the fluctuations.  If you have an interacting system of multiple components, and you go through systematically and increase synthesis for each of them, you should be able to spot the one(s) that contribute most to the fluctuations — when their synthesis goes up, the fluctuations throughout the system go down.  Once you have the culprit identified, you may be able to use the new quantitative methods being developed in many labs (including the Paulsson and Xie labs) to determine the relative contributions of gene expression, partitioning error and degradation to the variation.  Given the rapid progress of these methods, the authors say, it’s only a matter of time before these matters can be routinely settled.

Huh D, & Paulsson J (2010). Non-genetic heterogeneity from stochastic partitioning at cell division. Nature genetics PMID: 21186354