Final Report Summary - SELFREF (The Evolution of Evolvability: How Natural Selection Shapes Itself) The project has massively exceeded our expectations with regard to productivity; insights gained and follow up research generated. Greater than the four expected publications have been published and more are submitted. The project has allowed me to establish contacts with an eminent set of European partners with shared interests and complementary expertise to follow up on the radically new hypothesis proposed by Prof. Szathmary and myself during the Marie Curie Fellowship.The aim of the project was to understand the mechanisms by which evolutionary systems could arise, specifically to understand the origin of unlimited heredity (information transmission) in chemical systems. At an early stage we developed a novel approach to the origin of information transmission by defining the problem at a more general level, i.e. in terms of the origin of a physical symbol system (PSS). A physical symbol system is a physical system that contains arbitrary tokens, and explicit rules, i.e. entities that operate on tokens, that are themselves specified by tokens. Thus it is a self-referential system. The genetic system is a PSS, enzymes acting as explicit rules, and nucleotides acting as tokens. It seemed to us quite natural then to extend insights from evolutionary computation and evolutionary biology to neuronal systems in which the physical symbol system hypothesis was first proposed by Simon and Newell, and where after decades of work it is still not understood what the physical basis of neuronal symbols is. This led us to propose the neuronal replicator hypothesis (NRH); that selection at the level of the organism (human/primate) had evolved brains capable of sustaining replication of information. It followed naturally that if neuronal replication of symbolic information is possible, then evolutionary computation can take place in the brain to produce adaptive thought and action.The core of the project involved understanding how an informational system can arise in two very different physical systems. The first kind of informational system is the chemical system, notably networks of autocatalytic particles and binary string replicators based on Kauffman's autocatalytic sets. We applied group selection to such networks enclosed in compartments and found that a network could be selected to sustain increased amounts of information transmission. The second kind of system considered was neuronal systems. We proposed that a similar principle of "compartmentalization" could occur in the brain, where neuronal selection (based on Dopaminergic reward) is capable of selecting for particular neuronal circuits from a population of such circuits, also for circuits capable of sustaining high fidelity information replication. Unlike the proposals of Gerald Edelman (who has already used the term Neural Darwinism and Neuronal Group Selection), we have proposed true neuronal replicators. We proposed that the capacity for neuronal replication with unlimited heredity in human brains was necessary for the origin of language and generative creative thought. Principles learned from our study of the prebiotic origin of template replicators could be extended to the origin of neuronal replicators, notably the requirement for "compartment level selection" to shape an underlying informational system. We describe the two streams of work in more detail below.In the chemical domain of the origin of life, it remains a mystery how nucleotides could have arisen. Nucleotides are important because they have the special property that the polymers that they form can be template replicated irrespective of the order of nucleotides on the string. This allows unlimited symbolic (particulate) inheritance. Prior to template replication there was no microevolution, only attractor-based heredity. We critically considered several claims about the means by which attractor-based heredity could transform into template-based heredity, one of which was the notion that selforganization of autocatalytic sets of polymers (a la Kauffman) in a flow reactor could accumulate adaptations. In a paper in preparation for PLoS Computational Biology (ref) we have shown that catalytic networks can be shaped by group selection if they are enclosed in compartments. Importantly, they do not have the organization proposed by Kauffman, i.e. they are not typically reflexive autocatalytic sets; nevertheless they can undergo variation that allows them to transmit increasing amounts of information. In the absence of compartments (especially in the presence of inhibition) they cannot accumulate adaptations. We are working in collaboration with Staurt Kauffman on this publication. In a second submitted paper (ref) we show the conditions in which serial dilution in the famous Stanley Miller experiment of the origin of life can be expected to yield increasingly informational replicators. If the probability p that an autocatalyst produces another autocatalyst by a side-reaction is smoothly heritable then there will be a tendency for network growth to be biased towards regions of chemical space with higher p. This is an important finding because the proposed experiment is within reach, but has not been conducted so far, and the paper proposes a very general tendency of autocatalytic physical systems to tend towards systems capable of information transmission.Extending these insights to the origin of information processing in the brain, our neuronal replicator hypothesis proposes that replication of (a) digital synapses, (b) patterns of synapses (connectivity), (c) patterns of bi-stable activity, and (d) spatiotemporal spike patterns, takes place in brains. We discovered that Hebbian learning (a well known neuronal mechanism) is capable of structuring evolutionary search by learning the principle components of previously visited local optima. This allows solution of certain problems in polynomial time that were previously only possible in exponential time (ref). We have submitted a paper showing how neuronal replication solves the Stability-Plasticity Dilemma, whereby replication of actor-critic controllers can solve the problem of catastrophic forgetting in a simulated robotic task (ref). The implication is that neuronal copying may first have evolved for memory, and later have been exapted for artificial selection.The four proposed replication mechanisms are given below.(a) We have published a mathematical isomorphism between Hebbian learning and Eigen's replicator equations, thus providing an evolutionary basis for Adam's proposal of the quantal synapse (ref).(b) Copying of patterns of neuronal connectivity occurs by causal inference: Using Dr. Eugene Izhikevich's neural model, we published a paper in PLoS ONE showing how copying of patterns of connectivity (circuits) could occur by STDP based causal inference between layers connected by a topographic map. Generation times are in the order of minutes using Bonhoffer's recently discovered synaptic remodelling (ref).(c) Using neuromodulatory gating and bistable neurons, we showed how rapid replication of patterns of neuronal activity could be replicated using topographic maps, with generation times in the order of milliseconds.(d) Symbol systems can be implemented as spatiotemporal spike patterns: Using Izhikevich's neural model, we submitted a paper to Cognitive Science (ref) proposing how a physical symbol system could be implemented in the brain. This system for the first time proposes a neuroscientific basis for linguistic compositionality and systematicity, and the learning of general rules.In conclusion this Marie Curie Fellowship has allowed me to develop the neuronal replicator hypothesis that most unexpectedly takes insights from pre-biotic evolution and extends them to neuroscience. If the hypothesis is shown to be true, this work will be seen as a major advance in evolutionary theory and neuroscience.