Two weeks ago, an interesting commentary by Paul Nurse, came out in Nature.
The bottom line? We need to change how we study and understand cellular signaling cascades.
First, some background. Cellular function is governed by a network of protein interactions that act like an information processing device. These devices sense external inputs, such as cell signaling factors, pH, nutrient availability and temperature, and they regulate a vast number of different cellular responses such as changes in morphology, alterations in the metabolic state, the modulation of cell division, or the regulation of cellular differentiation. These information processing devices are composed of bioactive molecules such as proteins and functional RNAs.
Traditionally these devices have been known as signaling cascades. In most papers and text books they are represented by flow charts. Molecule X activates molecule Y, molecule Y inhibits molecule Z, etc. Often one uses the metaphor of a river so that molecule Y is downstream from molecule X and upstream of Z. Sometimes small molecules such as cAMP or calcium spread the signal, other times its lipids, RNA but the main players are usually proteins. Within the pathway each player activates, inhibits or destroys the next in line. Sometimes the upstream molecule alters the localization of the next protein in the pathway, other times it modifies these downstream components by adding a small reactive group (e.g. a phosphate), a small peptide (e.g. ubiquitin), or by modulating some catalytic activity (e.g. activating GTP hydrolysis). Once in a while a molecule at the end of the pathway will alter the activity of molecules on the top of the cascade. These feedback loops give these signaling pathways a rheostat like ability. Other times the pathway may branch (molecule X activates molecules A, B and C), and other times the pathways converge. Some pathways "crosstalk" to other pathways so that signals that affect cell division also tweak signals that influence cell suicide. In addition the levels of each player can change. Gene expression and protein turnover thus play critical roles in determining how effective a signaling cascade can propagate or inhibit a signal. As a result, different cell types vary in their interpretation of any given signal. It's why a gut epithelial cell differs from a kidney epithelial cell.
But is this the best way to think of these signaling cascades? With the advent of big biology, we are discovering that most signaling proteins can influence the state of almost every other protein in the cell. Once we had a clear picture of signaling, but now our view resembles more and more a bowl of spaghetti. Naturally we must question whether all these putative links exist and whether they are important.
From many studies, we can guess that the chaotic view of cell signaling is wrong. If you disrupt a gene involved in the integrin signaling cascade, your transgenic mouse will invariably have the exact same phenotype, death at a particular embryonic stage with a particular defect in the heart. If you don't get this phenotype, the gene you knocked out is not essential for integrin signaling. So yes, perhaps that protein can interact with integrin signaling molecules, but it is likely that the interaction is biologically irrelevant. You can perform other experiments where other complex cellular phenomena are measured, and almost always you will find that certain proteins are important for that particular phenotype, others aren't. Things may not be exactly black and white, but for the most part it is as close as you can get.
It does appear that most signaling cascades can be rewired easily throughout the course of evolution. This plasticity can show up in these huge in vitro protein-protein interaction data orgies, but may not give a clue as to how the pathway is hooked up in a given organism or in a given cellular context. In addition, similar proteins are used again and again for seemingly different purposes. In one cell a protein will play one role, in another cell it will play another role. It looks like chaos, but the protein has only certain effects, although this may vary from cell to cell, from organism to organism.
What's going on? The idea that Nurse is advancing is that we must move beyond these simplistic flow chart diagrams. In reality the molecular components of these signaling pathways form functional modules, each module resembling a component of an electrical circuit. These units are likely to be composed of several molecular players that interact tightly and in concert. The modules would have definite roles in how they affect how a signal is propagated within a cell. They could act as capacitors, amplifiers or some exotic device that has no known circuit-module analogue. Units can act together to form larger modules that have more sophisticated properties. These could act as signal integrators, positive- or negative-feedback loops and other useful signaling widgets. It is likely that the architecture of each unit is highly conserved (for example the MAP kinase cascade) but that the connections to other units are weak and malleable. Two units can be hooked up together or unhooked from each other depending on the cell type, or throughout the course of evolution. A byproduct of this loose linkage would be an increase in adaptability between different cell types and evolvability between different organisms. In contrast to the flowchart model, this vision of cell signalling as crazy-circuitry resembles more closely what we tend to observe, and it makes more sense evolutionarily. Pathways are likely to be haphazardly thrown together and disassembled over the course of evolution. But smaller modules that are of great use because they serve to amplify, dampen or modulate signals, are likely to be conserved. Why? Just like a hamer or a screw, each module is so useful and can be used in so many different cell signalling assemblies that it's function is selected for and maintained.
The bottom line is that we should rethink cell signaling within this new context. But to understand cellular circuitry we need a description of these modules, how they work and how they hookup together.
Now Sir Paul had expressed similar sentiments in a talk he gave last year here at Harvard Medical School. At the time I was uneasy about these "Systems Biology" approaches. My biggest concern was how to map such hypothetical circuits if we don't yet understand what might be the players involved? How do we study hypothetical "capacitors" when were not quite sure which proteins form this unit. Although Nurse did not have an answer then, and doesn't provide one in this latest essay, nonetheless I have changed my mind. I think that we can do it. But the way to tackle this problem is not to take a bottom up approach, where we accumulate all the small molecular measurements and reconstruct the network from scratch, but rather the top down approach. We need to act like reverse engineers. The idea is to break cellular phenomena into smaller functional units.
Say you want to study the signaling cascade that turns on pheromone response, the way to deconstruct this particular "circuit" is to give yeast various stimuli while varying a few variables at a time (for example the concentration of pheromone, or the duration of pheromone exposure). The yeast will then display various response curves and response times. We then just have to analyze how the population reacts and then model the system using as few algorithms as possible. We can then use these algorithms to predict the number of circuits needed to mimic this phenotype in response to the variables we've altered. We then can go back and try to alter the circuit(s) by altering the players, for example by introducing mutations into key signaling molecules. Thus we can tease apart how the architecture of the signaling cascade.
Now the keys to breaking the code are
1) Need to measure event on a cell-to-cell basis. Often getting a mean value of the population is not informative - measuring the extract generated from that population is even less informative. If every cell displays a sharp biphasic response curve, but every curve is shifted slightly, the cells will appear to slowly change state when in reality they're not.
2) We need to change one variable and measure one output in such a way that it can be informative. This is not easy. For example Alex van Oudenaarden (see this post) showed how one could vary osmolarity (input) and measure the localization of a sensor (output). The trick his group used was to subject the cells to pulses of high and then low salt and simultaneously measure the localization of fluorescently tagged Hog1, a protein that responds to the osmolarity signaling cascade. From there he was able to deduce the simplest differential equations necessary to model the input output response.
3) We need some knowledge of the molecular components. You don't only want to model your system but go in and tinker with it and remeasure it. Don't try to model something that you have no clue about - go and read the literature. This is the hardest lesson. Too many times I've encountered modelers who are clueless as to how the cells are organized, how they might respond to certain conditions and what might be the molecular components that make up each of these hypothetical modules.
So although many out there have looked down on Nurse's essay, I think that he's essentially right. In dealing with the problem of complexity in biological systems we may be able to gain much insight by looking at cell signaling from a different angle.
Maybe a "signaling cloud" would be a better analogy?
Neat, thanks for this. And it will be interesting to see how this sort of thinking is applied in the future.
this was already tried with the Alliance for Cellular Signalling over 5 years ago to much fanfare and headed by another Nobel lauret. It has since fizzled out. Perhaps these things are not yet doable
The idea that we can go from the bottom up has been useless. We've gained no insight from those approaches. What van Oudenaarden demonstrated is that we can study the process using a top down strategy and thus gain insight.
I still think that at the basis of these cascades there are proteins - small machines with definite functions. And that these machines come together to form distinct units such as the MAP kinase cascade, the TORC complex (see diagram above) or the 48S pre-initiation complex. But to get at what those functions are we need to start with the phenotype and work our way down not try to represent every interaction and hope that the tangled mess we create adds insight.
Ideas are cheap, but a proof of principle is priceless. The major advance is not Nurse's ideas (or the formation of some consortium), but the work from the van Oudenaarden Lab. That's why I pay little attention to philosophical arguments and do not agree with J. Wilikin's distaste of the use of "information" when discussing biological models. A model is only useful if it provides insight - van Oudenaarden's work demonstrates that this approach can lead to advances in our understanding of how these protein networks function.
One of the most useful anologies I found was to think of the signalling pathways like a flight map. Each protein in them is a destination, with lots of flight paths heading out from each one.
Hi Alex, sorry for being late in checking back for comments. You said though:
That's why I pay little attention to philosophical arguments and do not agree with J. Wilikin's distaste of the use of "information" when discussing biological models.
I'm sorry Alex, but you can't inject phrases like "information processing devices" and not get philosophical or anthropomorphic. Calling it something such as information processing only backs up to vague hand-waiving with a nice catch-phrase at best (if not totally wrong), when calling it what it is - an osmo-adaptation response - would be correct.
To the point, a search for the term "information" in Mettetal et al. turns up only three hits - one in the abstract, but as even serious researchers use meaningless catch-phrases or buzz words to describe their work. The other two use the word information not in reference to signal transduction, but to facts which the author knows of.
So that's an entire paper where the authors describe a theoretical model without using the buzz word. So much for the utility of using "information" in discussing biological models.
Eh... maybe I just get too annoyed at pointless buzz words though. There are very interesting aspects of the paper, and a new-enough understanding of biology to account for getting published in Science, but not so much to trump that of the many other biology papers published in Science.
Play back an audio recording, and you hear sequential events: you register a particular 'kind' of information, mediated by the design of your ear and brain. Now make a spectrum analysis of that recording and you see something new that you could not see before: a profiled facet of the dynamic components of that recording, where time has been integrated into different axes.
The presence of a spike somewhere, at some particular frequency, alerts us to a feature of the recording that was invisible to us when apprehending the recording sequentially, and therefore therefore represents something of a paradigm shift. It's the same information, but we can gain new insights into its behaviour by looking at it a different way.
Similarly, modern digital signal processing devices make use of z-transforms to allow us to view the information in those signals in a different manner, and 'play with' and manipulate the information in signals in ways that would be literally unimaginable before that kind of mathematics became useful in that field.
This is the kind of re-implementation of information processing that I think Nurse is talking about. We need to find a way to make some kind of 'spectrum' view of the pathways that are commonly found in nature, and they might tell us a lot of things that are... well, possibly unimaginable at this stage.
Or they might not - it's impossible to tell unless someone tries. Haven't mathematicians trod this path before, using graph theory for generalising the dynamics of flat network flow? That would be the first place I'd look if I were a mathematical biologist...