Dueling Selectively With Darwin
Kauffman, Stuart A.
“Dueling Selectively With Darwin,”
Personal Communication, The Scientist, 10 August l987.
(Kauffman is a professor of biochemistry and biophysics at the University of
Pennsylvania Medical School. He is also the editor of The Journal of Theoretical Biology, and a recent winner of a MacArthur Foundation “genius” grant.)
Turning points in my intellectual life have never been welcome; I always seem to resist them until forced to do otherwise. One such passage occurred some 10 years ago, as I was walking one spring morning in the Downs of southern England with the evolutionary biologist John Maynard Smith and his biologist wife Sheila. John, remarking on our proximity to Charles Darwin’s home, chided me gently: “You really must think about natural selection, Stuart.”
How his comment shocked me! Of course I should think more about it! But I had spent more than a decade exploring the idea that much biological order might reflect inherent self-organized properties of complex systems, even in the absence of selection. Since Darwin, of course, we have come to view natural selection, sifting out rare useful mutations from myriads of useless ones, as the sole source of order in biological systems.
But is this view correct? Might not complex systems spontaneously exhibit order? I had begun to ask this question in medical school, when Francois Jacob and Jacques Monod published their famous operon model. Biologists began thinking of the genome as a kind of biochemical computer, in which a gene or its products turn other genes on or off.
This view, first worked out in bacteria and viruses and now being extended to eukaryotes, implies that cell differentiation in development from the fertilized egg is mediated by a complex genetic regulatory network that coordinates synthetic activities of the roughly 100,000 genes in each cell type of a higher eukaryote. By current criteria, a mammal has on the order of 200 to 300 distinct cell types. The regulatory network is thought to control gene expression patterns in these different types.
To analyze the problem, it’s useful to simplify and imagine that each gene can only be active or inactive. Think of a genomic regulatory network as a computer with an on-off switch representing each gene’s activity. Since each gene can be on or off, there are 2/100,000 possible patterns of activity– a number large enough to catch the attention of even Carl Sagan.
How are we to understand a system with 100,000 genes switching one another on and off? In part, by our natural bent for reductionistic analysis. But even should we succeed in analyzing the detailed circuitry in the face of the genome’s scrambling the regulatory system in evolution, we need to integrate our knowledge and understand what features of that circuitry mediate the order we see.
At this point my old interest in self-organization suggested a new viewpoint. I had studied logic before medicine; hence the idea of “logical switching circuits” seemed a reasonable way to approach genomic networks. The question I posed early on was whether the richness of connectivity in a genomic network–that is, the number of genes that directly regulate any specific gene–might have an important bearing on the spontaneous emergence of orderly behavior in model genetic networks. To my delight, the answer was yes.
This fact still astonishes me. Consider a model regulatory system with, say, a mere 10,000 on-off genes. Hook the genes together randomly, with each gene directly regulated by only two other genes. Then assign to each gene at random one of 16 possible logical switching rules. Since such a network, which has both random “wiring diagrams” and random “logic,” is supposed to model a real genomic system, once it is constructed its structure is fixed.
It is therefore a random sample drawn from the pool of all model genetic
regulatory systems built with the same constraints on numbers of genes and numbers of inputs per gene. Do such random systems typically behave in an orderly fashion?
The surprising result I found over 20 years ago is that if each gene has only a few direct input genes, which is true in bacteria and viruses and may well be true in eukaryotes, then a system with 10,000 on-off genes settles down to one of only a few recurrent patterns of gene expression.
Those patterns are also stable to perturbation: If a gene’s activity is transiently reversed, the system typically returns to the same pattern. If we think of a recurrent pattern of gene activity as a cell type in the repertoire of the genomic system, then these “random networks” exhibit an order that is strikingly predictive of features seen in organisms. The cyclic patterns are stable to perturbation, mimicking the homeostatic stability of cell types.
If a cyclic pattern is a cell type,
the typical features of these networks is that any cell type can differentiate directly into only a few neighboring cell types, and from those to a few others. We know that the ontogeny in all higher eukaryotes takes place by just such sequences of branching patterns of cell differentiation. Is this due to natural selection? Or is it a universal feature of ontogeny despite selection?
I have wanted to believe that such deep properties of ontogeny as the prevalence of branching pathways of differentiation reflect the self-ordered properties of complex genomic systems, not selection.
More generally, the fact that randomly assembled model genomic systems exhibit marked order even roughly reminiscent of that found in organisms strikes a blow at our world view, in which selection is the sole source of order in biology. I think that view is wrong. Complex systems exhibit far more spontaneous order than we have supposed, an order evolutionary biology has ignored. But that realization only begins to state our problem, for Maynard Smith’s admonition is correct.
We must think about natural selection. Now the task becomes much more trying, for we must not only envision the self-ordered properties of complex systems but also try to understand how such self-ordering interacts with, guides and constrains natural selection. It’s worth noting that this problem has never been addressed.
The challenge has set me thinking about how selection interacts with such self-ordered properties. This job is hardly begun, but several points are clear. First, two kinds of “complexity catastrophes” tend to limit the capacity of selection to attain genomic regulatory systems that are extremely untypical in the ensemble of possible genomic systems. Classical population genetic results have long hinted at a limit to selection’s power to achieve “maximally fit” genotypes in the face of a constant mutation rate as the number of loci in the genomic system increases. Eventually, mutation overwhelms selection and disperses an adapting population away from optimal genotypes. But a second limitation on selection seems to be emerging.
Natural selection is a kind of combinatorial optimiation process. Typically such processes face a rugged, multipeaked “fitness landscape” due to conflicting design requirements. Under strong selection, a population will at least climb to a local peak. Simon Levin at Cornell and I found recently that as genetic networks under selection become more complex, attainable fitness peaks become lower! Worse, this appears to be a general tendency in any combinatorial optimization process.
As the entities under selection become more complex, the optima that can be reached become progressively more mediocre. Does this mean that even strong selection cannot achieve highly complex and precise systems?
Perhaps selection results in organisms that can adpat well because they “adapt on” fitness landscapes that escape these tendencies. What in the post-Darwin world this might imply has me even more deeply puzzled.
Dueling with Darwin? Not really. Embracing him, and moving on.