Current approaches to the neurosciences are naïve and often misguided. Contemporary researchers are hopelessly enthusiastic about computer simulations, wiring diagrams or connectomes, and brain activity maps. We may need software tools to visualize brains, but they will not provide any deep understanding of the brain itself.
I shall argue that the fundamental, discrete units of the nervous system are its mitochondria. The feature that we expect of an irreducible neural component is excitability. Mitochondria take excitability to an extreme. If mitochondria are the fundamental units of the nervous systems, then in any CAD model of the brain, they are precisely the parts to which the most care and attention should be applied.
Brains by Design
Well over a century ago, Santiago Ramón y Cajal sketched every observable feature of the nervous system. Despite astounding advances in genetics and cell biology, there is still no practical blueprint to guide the implantation of any type of hardware into the brain.1
In circuit engineering, netlists are connectivity diagrams illustrating the connections within or between individual chips. Once the logic of a circuit is verified and the complete netlist generated, simulations are used to determine how actual performance differs from the perfect connections of the netlist. In order to determine whether a netlist meets real world constraints, the circuit engineer uses a program called an autoroute. The program optimizes timing and interconnecting paths. It can also specify how much copper should be laid to meet the restrictions on capacitance and inductance imposed by high frequency signals, or the reverse. In the layout of a nervous system, much the same considerations apply.
CAD generally proceeds from the ground up. Completely defined one-dimensional sketches are stacked to form sub-assemblies and, ultimately, a top-level model. Until its scale and operation are fully constrained, the model is incomplete. Every feature of the model must be defined explicitly or by means of its relationship some other declared feature of the model. This is the whole point of the CAD modeling exercise: the ability to control global form and behavior by toggling a few critical design points.2
Coaxial Conundrums
Neuroscience research has come to provide more and better details about the brain. Marshalling this knowledge into CAD form is hardly a trivial task. Many conundrums are geometrical in nature. Why do neurons adopt a polarized axodendritic form? Why is their axon chiral? Which direction does myelin spiral, going down the axon, and across adjacent axons geared together in nerves, and among the many arms of a single oligodendrocyte? Why are Schwann cells single-use items while oligodendrocytes are multiplexed?3 How were whole axon tracts warped and decussated as the genetic body plan twisted and inverted during the evolution of vertebrates?4
Many signature functions of the brain have also been incorrectly or insufficiently explained. If the CAD model adds myelinating compartments to the axons, bundling them into a nerve fascicle or tract, how would they mate and slide together? A force drives the spiraling inner myelin strands. Does it derive from beyond the myelin itself, perhaps from torque supplied by the spiking axon inside? Do axons then wrap up their own myelin? Neurobiologists have no idea how myelin wraps, and they will have no realistic model until they do.5
Connectomes
Jeff Lichtman presented a 1-terabyte computer model of a tiny grain of mouse cortex at the annual meeting of the Society for Neuroscience in 2013.6 One particularly compelling reconstruction contained a cylindrical patch of tissue surrounding a single apical dendrite of a pyramidal cell. Lichtman’s team exhaustively mapped every mitochondrion, every postsynaptic density, and nearly every vesicle in the 774 synapses made on the dendrite by some 680 surrounding local nerve fibers. Using this approach, they extracted the complete membrane topology for one-billionth the volume of a mouse brain. The few advances made since have only served to drive home the obvious state of affairs, that it would take a great many terabytes to scale up to something the size of our brains. If we ever hope to have a connectome faithful to the living facts, we will have to be a bit more creative.
If researchers were serious about generating a connectome, they would not be talking about grey, but white matter.7 Neuroscientists focus on grey matter because they happen to be good at making detailed electron microscope (EM) sections; these provide a lot of information about tiny pieces of brain. Trying to map white matter pathways in this way would be like using a magnifying glass to see the Grand Canyon. If they are obviously too big to be seen using a magnifying glass, they are also too small to be resolved using mapping techniques such as diffusion MRI tractography.
That is a shame, because gray matter is highly labile. Aside from the obvious data problems, a faithful gray matter netlist is physically impossible. Neural connections change much faster than the time it would take to read them out. White matter, on the other hand, is both stable and predictable. Neighboring axons surrounding any given axon tend to signal in much the same direction, have much the same orientation, and are likely to be wrapped by arms from the same oligodendrocyte. In writing a white matter connectome, one might forsake a neuro-centric description altogether, and use instead the coordinates of each oligodendrocyte, together with the fifty or so of its minion axons.8
How plastic is the nervous system? How labile are neurons? Turning high resolution microscopes to the brain and looking for anomalies, researchers have found that the initial axon segment, where the axon roots itself at the cell and its organizing centriole, moves around the cell quite alarmingly. In up to 10% of cells in the hippocampus, the axon migrates from its normal location and roots itself on a proximal dendrite.9
A further consideration: is the brain made up of discrete parts? Or is it a syncytium, a multinucleated continuum? Cajal believed that nervous systems are made from discrete parts. Cajal was the first to declare that the neuron, with all its dendrites acting together to feed a single polar axon, is the fundamental unit of the nervous system. Camillo Golgi’s continuous syncytium model held sway before the work of Cajal.10
We know that organelles, such as mitochondria, are transferred from neuron to neuron. This presents a challenge to any doctrine based on discrete parts.11 If neurons routinely exchange structural wetware, then the differences between the old syncytial and the newer discrete models would seem significantly less sharp.
From a strictly mechanical point of view, the brain is a syncytium. While the electrical component of a spike that invades the synapse is chemically transformed into a vesicle message, the mechanical portion of the spike is transformed and propagated into and across the synapse in a way dictated by the mechanical impedance of the local membrane, cytoskeleton, and synaptic matrix.
Can We Construct Model Brains from Principles?
For several decades Eric Kandel’s Principles of Neural Science has been the standard textbook in neuroscience. Although the bulk of this text has remained fairly solid, improvements have been only incremental. In 2015, a second great tome was published, Peter Sterling and Simon Laughlin’s Principles of Neural Design.
Sterling and Laughlin offer a partial explanation for why things look the way they do in any given cube of tissue. Their analysis of different neural circuits suggests that there is some underlying logic to why neural connections have the precise diameters, numbers of vesicles, vesicle release probability, and spontaneous and maximum firing rates that they do. They espouse computing with chemistry, rather than with neural circuits, whenever possible, minimizing wire in converging and diverging circuits, and optimally locating cell bodies within the folded layers of various cortices and nuclei.
Their principles suffer from a narrowness of perspective. They are derived entirely from electrical considerations: how much energy it takes to generate an electrical spike in an axon, how much adenosine triphosphate (ATP) it takes to pump the ions back out, and how noisy the channels and receptors are. Their fundamental parameter is the electrical resistance of cytoplasm. However, the mechanical nature of pulse propagation in neurites, particularly in those where myelin participates in carrying some of the energy of the signal, suggests that sending spikes appreciable distances may be more efficient than had been previously assumed.
Although simulated networks attempt to connect and integrate signals, they invariably ignore certain important features of actual neurons. First, neurons sometimes connect to themselves. These autapses may represent a minor portion of their overall synaptic budget, but they are tightly controlled by a suite of recognition molecules. Second, when neurons do connect to another neuron, they do so with a bloom of synapses that defy characterization by the single synaptic weights used for artificial neural network connections. Cajal’s drawings in his Butterflies of the Soul show that retinal cells dedicate nearly their entire axonal or dendritic arbor to just one or two vertically adjacent partners.
Rather than bugs, I think these are design features. Perhaps they are even principles.
Karl Pribram proposed four F’s for evolutionary biology: feeding, fleeing, fighting, and fornicating. These are top-level behaviors, far above the crude functions, algorithms, or computations that neural modelers embed in their artificial networks. The network modeler seeks to build complex functions from the interactions of approximated neural units. Each real neuron already contains all of these complex behaviors. If the horizontal transfer of genetic material is part of copying and reproductive behavior, simple bacteria display all four evolutionary motivators as well.
Where is their brain?
Computations were presumably done with chemistry at the level of receptors and their associated signaling pathways. Complex nervous systems later materialized from this base. Their detailed network structures may have represented the spatial optimization of specific metabolic pathways.
Neuroscience has a lot to say about receptors and channels, but little about why neurons use the transmitters they do. Transmitter chemistry must be suited to whatever computational task needs to be done at a given synapse. Some transmitter molecules might be brute force irritants, and others, metabolic dead ends.
These transmitters are generally not shuttled whole through each junction; their parts and physical influence are transduced. In some circuits, these parts are directly rendered as enzymatic transformation and receptor activity in the synaptic cleft. Others require uptake of the largely complete molecule by the postsynaptic cell, or other glial hosts, in order to actuate their functional groups.
One thing many transmitters have in common: their metabolites control the metabolism of mitochondria. These organelles are invariably concentrated in close opposition to each other at active presynaptic and postsynaptic sites, giving one the direct impression that they are there to communicate.
Clues as to how cells use mitochondria, and conversely how mitochondria might use cells to communicate beyond their borders, are now emerging from studies of our immune and hematopoietic systems. White blood cells not only utilize mitochondria for the killing power of their oxidants; they relocate them to the plasma membrane, and then dangle them as immunogenic lures in the bloodstream.12 Ailing mitochondria expose some of their bacterial-style formylated peptides, cardiolipins, and DNA on their surfaces.
When the bloodstream is breached, thrombin and other factors transform formerly quiescent platelets, first into prickly amoeboids, and then into giant super platelet balls. They do this by modulating the base carrier frequency of calcium oscillations in their mitochondria, which, together with a sudden reversal of proton pumping in their ATPase machinery, triggers an irreversible cascade.13
In the brain, astrocytes have co-opted peripheral immune system actuators, such as glycoprotein CD38, in order to initiate the transfer of mitochondria from astrocytes to neurons to repair damage after a stroke.14 Conversely, the two main mitochondrial proteins implicated in disrupting mitochondrial dynamics in Parkinson’s disease, Pink1 and Parkin, are the key regulators in the peripheral adaptive immune system. The fact that mitochondria are involved with presenting antigens is astonishing.15
In each case, the underlying activators can be traced to base level self-excitations occurring deep within the cristae of mitochondria. These nonlinear mitoflashes are now known to be proton-triggered.16 Mitochondria cannot change mitoflash potentials or repeat them at a rate approaching that of neurons. Parametric imaging shows that mitoflashes follow up their local proton and membrane voltage flux with a predictable sequence of calcium and redox sparks. Mitoflashes and spontaneous oxidative bursts have been found to be accompanied by significant changes in mitochondrial shape. Although this mechanical aspect of the mitoflash takes place at a rate of only 0.6 per hour in a given mitochondrion, there are thousands, sometimes tens of thousands of mitochondria in each neuron.17 A large army of these mitochondria, even if they take a while to recompose themselves, should be able to influence the excitability of the neuron itself. The mitoflash could initiate neural spikes. Neural spikes could, in turn, initiate a mitoflash.
The action potentials of neurons are multi-physical. They include mechanical displacement, pressure, and the absorption and release of heat, events following their own predictable course in time and space. Oscilloscopes or drawings illustrating a wave of inrushing sodium ions and outrushing potassium ions can be misleading. When axons fire, there is no spatially-localized pulse; the whole axon depolarizes.
To the dismay of neurophysiologists, who have been using antidromic collision techniques to trace axonal connections, spikes travelling on axons in opposite directions can pass through each other. If these observations, originally discovered in worms, hold true more generally, then the neuroscience literature is riddled with spurious results. Persistent spikes would be reflective of mechanical soliton-like waves, rather than of the annihilating electrical spikes characteristic of the Hodgkin–Huxley equations.
Oops.18
The Brain is What, Exactly?
Modern genetics has traced the origins of mitochondria to the symbiosis of two different bacteria with complementary metabolisms. Details of the different theories vary, but the main idea is that as the genome of one type of bacteria (the mitochondrial precursor) was reduced, the genome of the other (the host) grew. The host became the nucleus of the first eukaryote.19 After eukaryogenesis, mitochondria turned their attention to driving cell differentiation and multicellularity. Ultimately they crafted the nervous system.
Sterling and Laughlin’s original cover art depicted an image of grey matter. What we see are mitochondria nestled inside a convoluted system of narrow tubes. These are passageways that the mitochondria themselves have constructed by the power of their own respiration.
We are looking at an elaborate ant farm. Ants moving on a typical ant trail do not migrate as a group. They run into each other, butting heads to exchange chemical status. They then change heading and repeat. From this simple back and forth, ants manage to create adaptive colonies. The commute itself is the computation.
This is more or less what mitochondria are doing. Mitochondria, but not ants, are capable of fusion. By fine-tuning their fusion and fission rates, each cell maintains a large centralized mitochondrial syncytium that dispatches and recalls mitochondrial quanta to different parts of the cell.
The spatial extent of this syncytial mitochondrial fluid both controls, and is in turn controlled by, the phase of the cycle. As most neurons are post-mitotic, their mitochondria are decoupled from the cell’s state, freeing them to drive their axon into parts unknown. Each mitochondrion is responsible for policing both its own health and that of the host cell. Fusion and fission provides a competitive mechanism to promote desirable mitochondrial DNA (mtDNA), purge oxidatively-damaged DNA, and ensure that the required host-cell proteins are properly distributed. Exactly why mitochondria form syncytia is not yet completely understood, but several ideas have been proposed, based on analogous aggregation of slime molds during times of stress, and from other schooling, flocking, or swarming behaviors.
It is not only mitochondria that routinely fuse, but whole cells, often quite predictably. For example, bone-resorbing osteoclasts, developing muscle cells, placental cells, fly embryos, and worm germ cells all form multinucleated syncytia. When foreign stem cells containing heteroplasmic mitochondria are introduced into the brain, they have the curious habit of fusing with the local population of neurons.
If whole neurons or their parts habitually fuse, a powerful new mechanism emerges to help explain how the brain could arise. Fusion anastomoses like those so common in the circulatory system could dissolve borders across pre- and postsynaptic sites to yoke neurons together. The creation of synapses in the middle of a bare neurite could isolate them again. Developmental curiosities and anatomical enigmas would no longer seem quite so inexplicable.20
More Mitochondria
Genetic-sequence analysis has tagged mitochondria as the direct descendants of a class of bacteria known as alphaproteobacteria.21 Still, there are many very different, very diverse bacteria in any given bacterial subgroup, and all of them have different metabolisms from present-day mitochondria.
Among geophysicists, the idea that the original proto-mitochondrion emerged from the class of magnetotactic bacteria is still current. The theory persists despite the fact that modern mitochondria show virtually no hint of having ever had magnetosomes or magnetotactic behavior.22
Precursors to the magnetite (Fe3O4) and greigite (Fe3S4) assemblies used to build magnetosomes appear to have originally played critical roles as terminal electron acceptors in various electron transport circuits. Whether catalytic FeS clusters came first, or even before the iconic tetrapyrrole cofactors so critical in every metabolism, may seem like remote origin-of-life questions. But such questions are critical in defining the instinctual behaviors and functions of mitochondria, which later guided their construction of nervous systems.
Intracellular and trans-cellular mitochondrial networks explain things about the nervous system that purely electrical considerations cannot. The elaborate dendritic trees of cerebellar Purkinje cells contain tens of thousands of synaptic inputs, all funneling down into the soma by means of single dendritic shafts. Sterling and Laughlin give a fine account of the firing rates, vesicle release probabilities, and the placement of different parts of the cerebellar circuit. What they do not do is explain the huge information loss if the Purkinje tree is regarded as an electrical machine.23 There is insufficient bandwidth available to convey all the messages to the soma, unless most of the dendrites are quiescent. Whatever the combinatorial logic occurring at the dendritic bifurcation points, it would appear that most of the bandwidth is throttled.
If signals are sparse, what an inefficient allocation of resources.
Similar arguments also apply to the axon. The same spike signal is sent to hundreds or thousands of synaptic endpoints. By any measure, this would seem to be a strange example of computer architecture.
If neuroscientists still are committed to the idea that brains and neurons execute computations, where might they occur? Is it all computing with chemistry, using receptors and intracellular signals? Could the electron transport chain have been harnessed by neurons in a way that parallels the refinement of electricity for modern electronics? Neurons are not apt to contain transistor arrays of memory chips or logic-gate chips. Nonetheless, the brain’s logical architecture may, in the end, be familiar to us. We do not, after all, have that many models for computation.
We cannot directly see the hardware of respiration when we look at EM images, but it is possible to infer much of what must be there. The folds and pits that concentrate and funnel metabolites between cristae take precisely mandated forms.24 The dimers forming the ATP synthase of the molecule called Complex IV are offset by 90 degrees. They bend the local membrane geometry to orient themselves in rows, at the bottom of a deep proton well.25 The positioning of the ATPase sets the floorplan for the three other major complexes involved in respiration.26
Spikes and synapses are both expensive in terms of energy requirements, but how does a given cell apportion the two? The probability that at least one vesicle is released after a spike is roughly fifty per cent throughout the brain. Does this merely reflect a balance struck between sending spikes and sending vesicles for each cell?
Not every synapse has a bulbous head poised on a restrictive neck, but this structure is common throughout the brain. Rapid mechanical twitches and shape changes have been attributed to synapse spines; no one has really explained their curious form. Not every spine has its own resident mitochondria. Active synapses nonetheless retain a captive power source in mitochondria. The swollen endbulbs of postsynaptic spines may well serve as their transient incubators.
Mitochondrial Bottlenecks
The informational bottlenecks from dendrites to axons are transformed in the brain, where they become physical bottlenecks with a new purpose: the selection and transmission of desirable mitochondria to the axon. I first suggested this idea a few years ago.27 Evidence for this type of process has just been found, along with several of the molecular mechanisms that underlie it.28 When combined with the known tendency of neurons to share mitochondria, this simple idea provides an explanation for the polarity of neurons themselves. Mitochondrial bottlenecks drove the transformation of un-polarized neuron-like cells, found in the primitive neural nets of jellyfish and hydra, into the highly polarized neurons and circuits we now find in the higher mammals. The evolutionary history can be seen in the progression from the ganglionic nervous system architectures of jellyfish up through all the other intermediary invertebrate creatures, and ultimately to us.
This taxonomy chronicles endless refinement.29
Traces of the primitive, more symmetric ganglia-style insect brains are all but gone in us, save perhaps for the unusual pseudo-unipolar dorsal root ganglion cells that still relay sensation up through our spines.30 The ultimate fate of mitochondria that manage to gain access to axons likely depends on the kind of circuit they are in. Some circuits may act as selective filters and transmitters. Another type is in evidence in the visual system, where mitochondria passing into the retinal ganglion cell axons are degraded at the optic nerve head, and then sent to specialized glial cells. They are then absorbed in a lysosomal fusion process that is akin to the resorption and turnover of spent photoreceptor outer segments. A strikingly similar mechanism is also responsible for elimination of paternal sperm mitochondria on fertilization.
The axon is not the only place where selective mitochondrial bottlenecks are found.31 A series of maternal bottlenecks can be found in the ovaries. They ensure that only the most desirable mitochondria are selected and transmitted by the nurse cells to the egg.32 In the germ cells of the hermaphroditic worm, cells fuse into a syncytium where the mitochondria can compete.
The amalgamation of polarized neural components into larger bottlenecking networks may provide a mechanism to explain even more esoteric phenomena. The transmission of acquired characteristics has always been thought impossible. This is no longer entirely true. One experiment involved training Planaria and then grinding them up and feeding them to other Planaria, who then went on to develop similar behavior, despite never having been themselves trained. These creatures possess a peculiar facility for regeneration through a population of adult stem cells, known as neoblasts, that are dispersed through their body.33
Extra-genetic transfers are difficult to establish, but demonstrations of Lamarckian inheritance in mice, and flies are equally difficult to ignore.34 They are also hard to explain by appealing to traditionally established epigenetic mechanisms. The paternal sperm bottleneck is bandwidth-limited. The sperm’s DNA also undergoes something like a genetic reboot, where the many slowly accumulated epigenetic marks on the DNA are wiped clean.
Some bivalves would seem to violate the rule of uniparental inheritance because sperm mitochondria evade degradation or extrusion in the egg. Contrary to appearances, this is not really a violation of a well-established evolutionary maxim. Male mitochondria are only transmitted from fathers to sons. The early embryo somehow distinguishes paternally-inherited mitochondria and sends them to one of the blastomeres known as 4D, which in males goes on to differentiate into sperm-making germ cells.35
Thinking and Breathing
According to the elementary logic typically applied to a sensory neuron, it receives and integrates information from the world and passes it to the next neuron. This does not always make good physiological sense. A significant portion of the information in any spike train, particularly during unstimulated spontaneous activity, represents what a neuron is telling itself. When neurons wish to communicate, they generally use vesicles and the occasional gap junction. Any external information that becomes directionally superimposed upon a regularly spiking cell only represents an incidental fraction of the information flowing throughout the neuron. Many spikes represent nothing more than the noise of the pump.
The primitive homeostatic function of a cell-wide or organelle-wide membrane potential originally served as a gradient to power membrane transport. Self-synchronizing spikes in a large neuron may have arisen so that endosymbionts might communicate. Any realistic model of a nervous system would contain responsive mitochondria to power the generation of its spikes.
Let us look at two extreme cases. The flukes of the blue whale are innervated by neurons whose axons are ten meters long. Spikes may be able to influence the entire neuron in real time, but there is no obvious way the nucleus can adapt its output to these constraints. Axonal transport of large organelles could take decades to reach their target, while even the fast pool, moving at a rate of several mm/day, would still take months. During development, the axons tethered to the growing whale’s tail are being pulled aft at the astounding rate of three centimeters per day. It is improbable that everything the growing neuron needs could be provided by its own nucleus. If the synapses in the tail can obtain fresh mitochondria and other supplies from local sources, their instantaneous energetic needs could be quickly met.
At the other extreme are fairy flies. These invertebrates are smaller than a paramecium. They achieve their compact form by jettisoning genetic and energetic machinery during development. By offloading the nuclei and mitochondria of a large percentage of their neurons, they reduce their nervous systems to the bare minimum. They live through extreme metabolic adjustments, and they do not live for long.
Although the development of all complex nervous systems requires the mitochondria, it now appears that their primary responsibility is not respiration. To discover the true nature of mitochondria, we need to look to those rare single cells that have discarded their own mitochondria. Metabolic processes typically grow around metal cofactors. The molybdenum cofactor clusters, iron-sulfur clusters and hemes, iodine thyroxins, cobalt cobalamins, and many others, have all played their roles in different niches of life. Mitochondria do their bit in constructing each and every one of these cofactors before handing them off to the cytoplasm. By toggling mitochondrial and cytoplasmic start sequences, cells can easily switch between synthesis sites.36
The various forms of mitochondria represent a compromise between the requirements of respiration and synthesis. Their synthetic skills are derived from their role in oxidation. The unique tubular phase of the cristae found in liver mitochondria is specialized to enable steroidogenesis; but the discoid cristae of sperm mitochondria are fused to form a large ringed syncytium optimized for ATP production. D’Arcy Wentworth Thompson’s quip that “the form of an object is a diagram of its forces,” may be particularly relevant to cristae. Not many models depict the structure of cristae, and even fewer try to capture their function.37
This is characteristic of the neurosciences today.
