Beyond reductionism – systems biology gets dynamic
Is biology just complicated physics? Can we understand living things as complex machines, with different parts dedicated to specific functions? Or can we finally move to investigating them as complex, integrative, and dynamic systems?
For many decades, mechanistic and reductionist approaches have dominated biology, for a number of compelling reasons. First, they seem more legitimately scientific than holistic alternatives – more precise, more rigorous, closer to the pure objectivity of physics. Second, they work, up to a point at least – they have given us powerful insights into the logic of biological systems, yielding new power to predict and manipulate. And third, they were all we had – studying entire systems was just too difficult. All of that is changing, as illustrated by a flurry of recent papers that are using new technology to revive some old theories and neglected philosophies.
The central method of biological reductionism is to use controlled manipulation of individual components to reveal their specific functions within cells or organisms, building up in the process a picture of the workings of the entire system. This approach has been the mainstay of genetics, biochemistry, cell biology, developmental biology, and even neuroscience. When faced with a system of mind-boggling complexity, it makes sense to approach it in this carefully defined, controlled manner. In any case, in most of these fields it was technically only possible to manipulate one or a few components at a time and only possible to measure their effects on one or a few components of the system.
The productivity of reductionist methods, and the lack of viable alternatives, brought with it a widespread but often tacit commitment to theoretical reductionism – the idea that the whole really is not much more than the sum of its parts. Appeals to holism seem to many biologists not just out of reach technically, but somehow vague, fuzzy, and unscientific. We are trained to break a system down to its component parts, to assign each of them a function, and to recompose the systems and subsystems of organisms and cells in an isolated, linear fashion.
We can see this in genetics, with the isolation of a gene for this or a gene for that. Or in signal transduction, with the definition of linear pathways from transmembrane receptors, through multiple cytoplasmic relays, to some internal effectors. Or in neuroscience, with the assignment of specific and isolated functions to various brain regions, based on lesion studies or activation in fMRI experiments.
The trouble is that is not how cells and organisms work. Defining all these isolated functions and linear pathways has been productive, but only from a certain perspective and only up to a point. This enterprise has mostly depended on analysing responses to strong experimental manipulations – a trusted method to perturb the system but one that is inherently artificial (what Francis Bacon, the so-called father of empiricism, called “vexing nature”)*. And it has mostly analysed effects on limited, pre-defined readouts.
That’s all fine – strong manipulations that sensitise a system can reveal roles of specific components in some process that would otherwise be undetectable. The problem comes in forgetting how artificial these scenarios are and in inferring greater specificity and isolation than really exist.
The failure of so many drug trials highlights the hubris of thinking that reductionist manipulations and analyses reveal biological systems as they really are. We tend to focus on the standout successes, as validation of the whole enterprise, but the hit rate is tiny, even for candidates with exhaustive pre-clinical support.
Systems in the wild
Real biological systems – in the wild, as it were – simply don’t behave as they do under controlled lab conditions that isolate component pathways. They behave as systems – complex, dynamic, integrative systems. They are not simple stimulus-response machines. They do not passively process and propagate signals from the environment and react to them. They are autopoietic, homeostatic systems, creating and maintaining themselves, accommodating to incoming information in the context of their own internal states, which in turn reflect their history and experiences, over seconds, minutes, hours, days, years, and which even reflect the histories of their ancestors through the effects of natural selection.
Living things do things – they are proactive agents, not merely reactive systems. This is widely acknowledged in a general sort of way, but the problem is that the conceptualisation of cells or organisms as proactive systems has remained largely philosophical, even metaphorical, and disconnected from experimental study. It’s just not easy to study cells or organisms as systems.
Systems biology was intended to take a Big Data-driven holistic approach, but until recently, had failed to fully divest itself from static, reductionist thinking and embrace an enactive, dynamical systems framework.
Big data alone do not solve the problem, without some dynamical systems theoretical underpinning to how they are analysed. In genomics, for example (or transcriptomics, proteomics, methylomics, whatever omics you’re having yourself), new technologies allowed researchers to assess the state of vast numbers of molecular entities, under various conditions. But these approaches simply produced ranked lists of thousands of genes or transcripts or proteins and then the question was: what to do with them?
One thing is to look for some kinds of patterns or cross-correlations in the data, which can serve to implicate various genes or proteins in related pathways or processes, through guilt by association. This can certainly be informative, but has its limits – it doesn’t tell us how a gene or protein functions in any given process nor how that process actually works.
A complementary approach is to focus on specific molecules for experimental follow-up using reductionist methods. Faced with such extensive lists, the chosen molecules were typically the ones already implicated by some prior evidence (the “go down the list and pick the one you would have studied anyway” approach), creating a rather circular chain of logic and not necessarily deepening understanding of the system as a whole.
Network analyses promised a more holistic perspective, but these have tended to remain very much at the descriptive level. Given a list of molecular entities under several conditions, it is possible to generate network graphs based on their pairwise cross-correlations or some other statistical interaction. By themselves, these analyses simply cough up colourful hairball diagrams.
However, it is possible to analyse these network graphs to define many different kinds of parameters of local and global connectivity, such as degree of modularity, efficiency of information transfer, and so on. In my view, these sorts of structural descriptors are not particularly illuminating. It’s not really clear what you can do with those figures, except compare them, but then the major message seems to be that most networks tend to have broadly similar properties (small-world connectivity, a few well connected hubs, etc.).
What is missing from these analyses is any sort of dynamics or an idea of what the system is doing. A step in the right direction is the definition and recognition of functional motifs. Elements in biological networks (such as genes or proteins or neurons) are not just connected any old way. The connections often have a sign – they may be activating or repressing, excitatory or inhibitory. And they also have an architecture – they may be feedforward or feedback or autoregulatory. If you connect elements of different signs in different mini-architectures, you get functional motifs – little units that can perform certain types of operations. (Well, most of the time you get nothing – only a small subset of all possible combinations is actually functional, and an even smaller subset is robustly so).
These kinds of functional motifs can be recognised in an abstract sense, in whatever kind of network you are looking at. They can act as filters, amplifiers, switches, coincidence detectors, evidence accumulators, error predictors, oscillators, and so on. By analysing the structure and nature of interactions between elements of a network (like transistors on a chip) it is possible to identify such motifs and infer something about what kind of operations that little subsystem can perform. And, of course, you can build more complicated systems out of these functional units.
|Reproduced from this really nice presentation by Kimberly Glass|
These structures don’t just pop out of cross-correlation data, however. Usually it requires more defined reductionist experiments to supply the necessary data. In addition, while such analyses promise some insight into what bits of a system can do, in the abstract, they don’t show you what the system, as a whole, actually does – either how it behaves, as a collective, or what that system behaviour correlates with.
To do that, we would ideally like to track the state of all the elements of a system over time (how much each gene is being expressed or how much each neuron is active, for example), correlate that state with some external measure (like cellular phenotypes, or organismal behaviour), and deduce some global functional relationships.
Revealing the dynamics
This presents both technical and computational challenges, as well as a deeper philosophical challenge – what is the right way to think about how biological systems work? This is where a number of recent studies have proved so exciting – to me at least – as they have tackled the technical and computational challenges with incredible ingenuity, and, in the process, are helping to give some concreteness and experimental testability to the enactive, dynamical systems perspective and a philosophical approach rooted in process rather than substance.
Impressive new technologies are meeting the challenge of collecting the kinds of data we need. For example, single-cell RNA sequencing can give a profile of the expression level of all the 20,000 or so genes across large numbers of individual cells in a sample, under various conditions or at various time-points during differentiation. Similarly, advances in genetically encoded calcium or voltage indicators and miniscopes (or other recording approaches) allow the recording of neural activity patterns across large numbers of neurons in awake, behaving animals. In both these types of scenarios, a huge amount of data is generated that effectively captures the state of the system across time.
The problem is those data are vastly high-dimensional. Each snapshot is a matrix of the present state of N elements, with the state of each element potentially changing at each measured time-point. The challenge is to identify the meaningful patterns within those high-dimensional data.
There are many possible approaches to this problem but one that has emerged recently, across diverse fields, involves mapping these data to a low-dimensional “manifold”. In essence, this relies on a kind of principal components analysis – extracting statistical patterns that capture much of the important variance that is correlated with some external parameter.
|From Saxena and Cunningham, 2019|
This is only possible because there is actually low-dimensional structure in the state of the system. Most complex dynamical systems, characterised by a network of simultaneously acting feedback interactions, can only stably exist in a tiny fraction of all the mathematically possible states of the system – so-called attractor states.
If we think of a system with just two elements, A and B, we can imagine what will happen if A represses B and vice versa, and each of them activates itself. The system can exist in a state of high A and low B, or low A and high B, but not any other states. Now expand that thinking, but to many thousands of elements, with many thousands of interacting feedback loops. These interactions place massive constraints on the possible states of the system, and, additionally, on the dynamics of the system – how it tends to transition from one state to the next.
Reducing the dimensionality of the data can thus reveal the underlying dynamics of the system as it moves from one state to another. If these states correspond to something – if they mean something – then this approach can illuminate not just the myriad details of what is happening in the system, but allow us to make sense of what it is doing.
Attractor states and dynamic trajectories in neuronal networks
In neuroscience, this approach was championed by Walter Freeman in the 1960-1980’s, inspired by his work in the olfactory system. He and his colleagues recorded the responses to various odorants in many neurons at a time in the olfactory bulb of rodents. They discovered that the responses of individual cells were noisy and unreliable, but the responses of the system as a whole had a discernible structure and dynamics, in that the odorants could be deduced by the experimenters from the dynamic trajectory of the patterns of neuronal activation through system space over some short time period.
|Taken from presentation by Walter Freeman|
Subsequent work by Gilles Laurent and others reinforced this idea. It has always seemed to me to be a powerful and insightful method to describe what a neural system is doing and understand what kind of information it is encoding and what sorts of things it cares about. But it never replaced the prevailing reductionist paradigm in neuroscience of linear, hierarchical signal processing, and I have rarely seen it extended beyond simple sensory systems.
That is definitely changing, with some recent compelling examples enabled by new technologies that allow researchers to record from very large numbers of neurons at once in awake, behaving animals, and new computational methods to analyse the data, including machine learning approaches to extract the meaningful, low-dimensional patterns.
These include, for example, characterisation of the attractor dynamics of head-direction cells in rodents, the flow and modulation of human cognitive states and Bayesian inference in monkeys. These approaches are nicely reviewed here and here, and harken back to Freeman’s seminal work, described in his book: How Brains Make Up Their Minds (2001, Columbia University Press.
The resurrection of Waddington’s landscape
Similarly, in the study of development, new technologies are reviving some old concepts. In the 1950’s, Conrad Waddington introduced the ‘epigenetic landscape’ as a visual metaphor to help understand the transitions in cellular differentiation during development of an organism. (And no, it’s not that kind of epigenetics). This metaphor depicted a ball rolling down a landscape with a series of forking valleys, or channels. The ball represented a cell and each of the channels represented a possible cell ‘fate’ that it could differentiate into, while the ridges between them represent unstable states. The ball can also represent the whole organism, with the various channels representing different phenotypic states.
|From "Innate" (Kevin Mitchell, Princeton University Press, 2018)|
The landscape is shaped by all of the cross-regulatory interactions between all of the genes in the organism. The contours of this landscape will vary from individual to individual due to genetic variation. As a result, the developing organism may be more likely to be channelled down one pathway or another in one individual versus another, with the outcome in any specific case being influenced by noise in the system or external forces. This can help explain the probabilistic inheritance of complex diseases, where differential risk can be inherited, but the actual outcome is not genetically determined.
Sui Huang provided a nice update and partial mathematical formalisation of Waddington’s framework in an insightful paper from 2011, but there have been few rigorous experimental studies that have fleshed out this approach.
A new study by Thomas Norman and colleagues, from the lab of Jonathan Weissman provides an impressive experimental test of this framework. They used a CRISPR-based technique to over-express hundreds of regulatory genes, in pairwise combinations, in mammalian tissue culture cells, and analysed their effects on the transcriptional state of all of the genes in the genome using single-cell RNA sequencing during growth and differentiation. They then used computational techniques to extract a low-dimensional manifold that captured much of the variation in the landscape of genetic interactions across all these manipulations.
This genetic interaction manifold revealed convergence onto a small set of biological processes, perturbation of which had definable effects on the transcriptional profile and differentiation trajectories of the cells. It highlighted the functional nature of genetic interactions and suggested new ones, which were experimentally verified.
This paper thus puts some experimental meat on the conceptual bones of Waddington’s landscape. In particular, it allowed the researchers to measure how different genetic variants can shape that landscape, singly and in combination.
Crucially, it is not just an analysis of the transcriptional states underlying different cell types (i.e., which genes are turned on or off in the formation of a muscle cell or a skin cell or a nerve cell). It describes how those developmental pathways are affected by genetic perturbations. In this case it was an experimental over-expression of specific genes, but the same logic applies to the effects of naturally occurring genetic variations, whether they increase or decrease gene function.
This sort of approach thus provides some hope of finally getting a handle on the high-dimensional, non-linear (or epistatic) genetic interactions that underlie the relationship between whole genotypes and phenotypes. The phenotypic effects of any particular genetic variant are massively constrained by all the gene regulatory interactions encoded in the genome (which are evolved to robustly channel development into certain outcomes) and by the simultaneous effects of all the other genetic variants present in the genome. These can collectively push the system into new territory and even reveal novel attractor states in phenotypic space, some of which may be pathological.
Up till now we have only been able to get a statistical or averaged view of these genetic effects. The computational approach of defining low-dimensional manifolds may allow us to understand the outcome of all these effects by following what are likely to be highly prescribed trajectories through phenotypic state space.
A philosophical shift
All of these examples suggest we are watching a paradigm shift in real time: systems biology is finally embracing dynamics and, in the process, providing the means to turn vast quantities of data into real knowledge, deepening our understanding of what living systems are doing. The success of these findings should help to rehabilitate holistic, enactive approaches in the minds of scientists wedded to what they see as the rigour of reductionism.
Enactivism sees organisms as creating their own reality through dynamic interaction with their environment, assimilating information about the outside world into their own ongoing dynamics, not in a reflexive way, but through active inference, such that the main patterns of activity remain driven by the system itself. This perspective is well described by Varela, Thompson and Rosch, and developed by Evan Thompson in his 2007 book Mind in Life, and by others, including Alicia Juarrero (Dynamics inAction) and Andy Clark (Surfing Uncertainty), for example. But outside of philosophical circles (especially philosophy of mind and related areas of cognitive science) these ideas have not had the revolutionary impact one might have expected.
Biology textbooks do not present this view of life and biologists rarely teach it to students. The reasons can perhaps be traced back to the adoption of the mechanistic perspective in Western science, following the successes of Newtonian mechanics, and the accompanying tradition of reductionism. This can be set against a very different conceptual tradition of process philosophy, as championed by Alfred North Whitehead in the early 1900’s. (Science and the Modern World).
Process philosophy offers a radically different conception of living things, one that is not mechanistic or reductionist or rooted in fixed entities or substance. Instead, it is emergent and holistic and rooted in continual flux and change – what is now called dynamics.
Both process philosophy and enactivism have always had some links to Buddhist philosophies, which (I think at least) have tainted them with an aura of mysticism and woo. That is unfair, as the Buddhist philosophies in question can be separated from any aspects of spirituality or religion (see here for a relevant discussion that I had the pleasure of taking part in). But to many scientists, the ideas of process philosophy (if they have heard of them at all) have remained too vague and nebulous to be considered truly scientific or to give any kind of experimental purchase.
The new approaches described above will, I hope, help to ground these philosophical approaches in experimental science, making them rigorous and quantitative, and hopefully demonstrating their power to help us, like Neo, see the meaningful patterns in the Matrix.
*thanks to Jag Bhalla for the reference and many other helpful suggestions