Scalable Neuroscience and the Brain Activity Mapping Project

 Abstract 

Since the beginning of the year, the European Union and United States have separately announced major initiatives in brain science. The latter is called the Brain Activity Mapping (BAM) Project² and the size of the effort and the implications for science and medicine have been compared to the Human Genome Project. A key part of the effort involves developing new scientific instruments capable of observing the activity of large ensembles of neurons in awake behaving humans with the goal of understanding the neural basis for cognition and diagnosing a wide range brain disorders from Parkinson’s to Alzheimer’s.

The problem these instruments are intended to solve can be divided conceptually into two parts: recording and reporting. Recording involves sensing and coding for transmission neural activity including membrane potentials, protein expression levels, calcium concentrations and their correlates. Reporting involves conveying the coded information from the locus of the recording — typically deep within the neural tissue of an awake subject — to some external computing or storage device.

The technical challenge involved in building these instruments is considerable, perhaps on a par with constructing the Large Hadron Collider (LHC), but while the LHC accelerator ring is 27 kilometers in circumference, the components comprising BAM instruments may include billions of nanoscale parts and be contained entirely within a human skull. This lecture explores several of the key technologies being considered to address the reporting problem including nanoscale communication networks, micron-diameter fiber-optic cables, light and ultrasound microscopy, recombinant DNA and synthetic biology.

“Das Leben, das Universum und der ganze Rest”, is German for “life, the universe and everything”. Posed as a question, it sounds much less mundane in German, and, according to Douglas Adam’s The Hitchhiker’s Guide to the Galaxy, it is the ultimate question. It would seem that Adams subscribed to the common opinion that, however complicated life may be, the universe is more so, and there’s something even larger referred to here as “everything” that is more complicated still. I would reverse both the implied order of complexity and difficulty of obtaining a satisfying answer, in lieu of having access to Deep Thought³ or the portion of that galaxy-spanning computational matrix which is supposedly stored in Arthur Dent’s brain⁴, and I would add the human brain as the ultimate among ultimate questions.

There are plenty of other scientists and technologists making similar claims and in recent months we’ve also seen some major players on the international science funding scene jumping on the bandwagon:

• The European Commission will award a total of two billion euros to two projects one of which focuses on basic research on the brain and brain diseases such as depression, Parkinson’s disease and Alzheimer’s and possibly develop new treatments. (Source: Reuters, Brussels, Monday, January 28, 2013)

• The Obama administration is planning a decade-long scientific effort to examine the workings of the human brain and build a comprehensive map of its activity, seeking to do for the brain what the Human Genome Project did for genetics. (Source: New York Times, New York, February 17, 2013)

In this lecture, we’ll investigate some new ideas and emerging technologies that have fueled the enthusiasm and optimism leading up to these two grand-challenges-in-brain-science initiatives. In physics, it often seems like the theorists have an edge over experimentalists in terms of numbers of Nobel prizes awarded, but that may have more to do with the fact that we tend to hear more about theorists in the popular press: Einstein, Feynman, Dirac, Bohr, etc. In any case, a theory without experimental confirmation is an unlikely candidate for a Nobel, and often such confirmation would not have been possible without the development of new scientific instruments surmounting considerable scientific and technological challenges, e.g., the Hubble telescope and Large Hadron Collider.

Without an appropriate instrument, it may be difficult if not impossible to infer the hidden causes that give rise to a given observable phenomenon. This is certainly true when it comes to understanding the brain. We are astute observers of human behavior but it has been extraordinarily difficult to trace behavior back to the neural circuits and biological mechanisms that caused it. We can see in some detail the activity of a few neurons or view as though through a cloudy glass the aggregate behavior of large collections neurons. However, to make the next steps forward, we need to track millions if not billions of neurons, simultaneously recording activity at many levels of detail. To do so we will need to automate much of what was previously done by skilled scientists and exploit the exponentially increasing returns provided by modern computational technologies to make sense of the resulting deluge of data.

Medical illustrators and animators are getting much better, if at times a little overly creative, in rendering stunning images of neural networks.

Our understanding of individual neurons has improved enormously in the last couple of decades due to new electron-microscopy imaging technologies. However, our grasp of how networks of neurons work together at the cellular level to carry out complex cognitive tasks is still rather primitive. By invoking Rutherford’s model of atom and this “Tinkertoy” neural network, I am perhaps giving our current understanding of brain-scale neural networks too much credit.

Nowadays most high-school students know what a neuron is and can answer basic questions on the AP exam concerning the function of dendrites and axons.

We now realize that neurons are complicated machines that carry out subtle computations we are just beginning to understand, and that signal one another using an incredible array of chemical, electrical, genomic and mechanical means. Hardly a month goes by when we don’t read about a new neurotransmitter, genetic pathway, or information-processing capability that has escaped our notice until now. And increasingly these revelations are made possible by new scientific instruments that allow us to better see both the inner workings of individual neurons and the collective behavior of ever larger ensembles of neurons.

Among the many different types of brain activity one might want to record, action potentials or spikes are high on most neuroscientists wish list. An action potential corresponds to an abrupt change in the electrical potential across the cell membrane initiated in the axon hillock near the neuron cell body or soma and propagated along the axon to the synaptic terminals. (The full-length version of the action-potential animation shown in the talk is available here.)

Once the membrane potential reaches the threshold for spike initiation, a spike is likely to soon follow and propagate along the axon to the synapses. If we could locate a nanoscale sensor in the axon hillock to record the local membrane potential that would be quite useful. In fact, it would probably be almost as useful to simply record when the potential exceeds the threshold.

There are other types of activity correlated with spikes that may turn out to be easier to record. When a signal is transmitted across a synapse from one neuron to another, calcium ions play a key role in the release of small packets or vesicles containing neurotransmitters. By recording the concentration of calcium ions in the synapse, we obtain a lagging indicator for synaptic transmission and thus the propagation of information from cell to another. (The full-length version of the synaptic-transmission animation shown in the talk is available here.)

We have the technology for recording membrane potentials and calcium concentrations, but reading off these signals at scale is potentially problematic. The sensors are essentially biomolecules that fluoresce when certain local states obtain. The patterns of fluorescence are read off by a very sensitive optical imaging device called a two-photon excitation microscope.

However, a two-photon microscope relies on light and, while it is able to penetrate deeper than conventional light microscopes, the depth of penetration for state-of-the-art technology is limited to at most a few millimeters. BAM researchers would like to read off such information for millions of neurons simultaneously at millisecond resolution in an awake behaving human subject.

Here is an image of a fly’s head obtained using a scanning electron microscope. The image file is large by consumer-camera standards and offers a relatively superficial view of the fly. The adult fly is 8-12mm in length and its brain is around a cubic millimeter in volume. A 3-D image of its brain obtained from a serial block-face scanning electron microscope requires about a petabyte of storage — that’s more than a million gigabytes or a thousand terabytes. The Max Planck Institute for Biology in Tübingen is the premier institution in the world for studying flies and in particular the fly brain. A mouse brain is about 10³ cubic millimeters and would require around a million terabyte disk drives.

In 1959 Richard Feynman, gave a talk at Caltech entitled “There’s Plenty of Room at the Bottom” in which he imagined technologies that would enable us to peer at individual atoms and — perhaps more compelling to tool-using Homo Sapiens — actually manipulate atoms.

Feynman anticipated the atomic force microscope and the enormous potential inherent in the ability to explore the world at the atomic scale⁵. He also anticipated the development of nanotechnology capable of interacting directly with atoms and molecules and building nanoscale machines. He was particularly enamored of these “tiny machines” as he called them in a subsequent lecture and he issued several challenges for the creation of particular machines offering prize money as an incentive to aspiring inventors.

Along with John von Neumann, his former colleague at Los Alamos, Feynman fully appreciated that nature had already solved the problem of atomic-scale machines, and he considered biological machines a proof that such technology was possible and only a matter of time before engineers would match or surpass natural selection⁶. He provided insights into the challenges faced by natural and man-made tiny machines, describing how, as you descend to smaller and smaller scales, different physical laws dominate. E. Coli use corkscrew-shaped flagella and molecule-sized motors to propel themselves through a watery fluid which is for them a viscous medium as thick as molasses. For organisms operating at millimeter scales, surface tension is an important consideration; at nanometer scales, Van der Waals force starts to play a key role⁷.

Today, we frequently see articles in the main-stream press describing advances in microelectromechanical systems or MEMS which are devices typically manufactured using semiconductor fabrication technologies and consisting of components from 1 to 100 micrometres is size. At this scale electrostatic surface effects dominate over volume effects such as inertia or thermal mass. We are also seeing new materials that are hybrids combining, for example, biologically-based substrates constructed by folding strands of DNA into three-dimensional shapes, and then adding atoms of gold or other exotic materials as conductors to implement specialized sensors and communication devices.

In 1965, Gordon Moore made the observation that over the history of modern computing hardware, beginning with the invention of the integrated circuit in 1958, the number of transistors on an integrated circuit doubled approximately every two years. This exponential trend has continued more or less unabated to this day and promises to continue for some indeterminate time into the future. In this graph, the fabrication process — a modern version of lithography — referred to here as “technology nodes” — is now around 22 nanometers which allows printed lines etched on a silicon die — referred to here as “gate lengths” — of around 30 nanometers. The molecular machines in your cells — called “ribosomes” — responsible for manufacturing proteins are about 20 nanometers end to end and ribosomes are considerably more complicated machines than a single logic gate.

This next graph is a little too busy for my taste but if you take the time to parse it you’ll learn some interesting facts. For instance, the size of a transistor in an Intel 8008, a microprocessor introduced in 1972, is depicted by the large bluish-purple circle and is about twice the size of a red-blood cell, and the size of a transistor in a present-day Xeon server is about half the size of an HIV virus.

The challenge of scalable neuroscience is to build instruments that enable us to record the behavior of ensembles of billions of neurons at millisecond temporal resolutions where each neuron is a machine of incredible complexity, and infer from this virtual deluge of data — “tsunami” is perhaps a more apt metaphor, the function of individual neurons and predict the collective behavior of an entire brain in both its normal and pathological operating regimes.

Much of modern experimental neuroscience is based on single-cell recordings of individual neurons or multiple neurons within a small, roughly planar area of brain tissue using an array of probes arranged in a regular grid — 10 × 10 is common — that is inserted into the brain of an awake animal. Ed Boyden and his team at MIT are pushing the state of the art to enable each probe in such an array to record at multiple sites along its length thereby allowing us to collect information from many neurons in a 3-D volume.

Ed has pioneered methods for using robots to insert probes in experimental animals thus eliminating one source of human error and allowing precise placement under program control [31]. He is also applying optogenetic techniques — which we will discuss in a moment — that allow us to use light to both activate and silence individual neurons.

As long as we have had microscopes powerful enough to resolve individual neurons, scientists have been refining methods for imaging neural tissues using specialized preparations that make neuron cell bodies stand out and utilizing ever more powerful devices, with scanning electron microscopes currently now common in academic labs. Once the tissue is prepared and an image taken, it is generally the task of a trained neurophysiologist to interpret the image and determine where one cell leaves off and another one begins. Having skilled humans in the loop, whether working with the tissue samples or interpreting images doesn’t scale, and so research labs led by Winfried Denk at Max Planck and Sebastian Seung at MIT are developing robotic devices for handling the tissue and interpreting the results of imaging [5, 29].

Unfortunately, automating the segmentation of cell bodies is more difficult than you might imagine [46, 38]. You can plainly see the leopard cub in the top sequence of images of this slide, differentiating its torso from the tree to which it clings. Segmenting the dendrites and axons in the middle row of frames is much more difficult. You may imagine individual neurons gracefully spread out in the neural tissue like free-floating seaweed fronds, but it is more accurate to imagine the neurons as spaghetti noodles densely packed into a can. The task is made somewhat more tractable by highlighting selected neurons using color-coded fluorescent markers [23] — see Brainbow — as shown in the bottom panel, but this technology is not likely to scale due to the combinatorics involved in differentiating so many closely packed cell bodies.

Sebastian’s goal is to compute the connectome — the graph of neurons and their active connections — for interesting tissue samples, starting with the retina, then a mouse brain and ultimately a human brain [45]. Even if we can improve our image processing algorithms to accurately segment cell bodies, we would still need a warehouse full of robotic tissue handlers and electron microscopes to process even a single mouse brain in a reasonable amount of time. A single cubic millimeter of neural tissue produces a petabyte of image data when scanned. One would hope there’s a better way.

The method of preparing a tissue sample, slicing it into thin sections, and scanning each slice with an electron microscope that is being used to reconstruct the connectome can also be applied to determine where in the cell different proteins are utilized. Stephen Smith and his colleagues at Stanford have developed a new imaging technique they call array tomography that combines electron microscopy with immunofluorescence to visualize the distribution of specific proteins in the cell [37]. Immunofluorescence takes advantage of the specificity of antibodies to their corresponding antigens to tag proteins with fluorescent dyes so they can be imaged with a scanning electron microscope. Smith and his team have used this technique to investigate the diversity of different synapse types as identified by their characteristic protein signatures [40] and the Allen Institute for Brain Science has used similar techniques in generating data for their incredibly useful Brain Atlas resources.

Let’s consider the protocol involved in array-tomography immunofluorescence imaging, a technique developed by Stephen Smith and his colleagues at Stanford [37]. As in the case of the connectomics protocol, we begin by perfusing the tissue with a fixative and preparing it for serial section. There are a variety of techniques used to cut the tissue sample into serial sections in an efficient assembly-line production manner. Once the serial sections have been prepared they are bonded to a coverslip to facilitate scanning electron microscopy. The next step is to “stain” the tissue bonded to the coverslip using one or more composite molecules referred to as “dyes” consisting of an antibody bonded to a fluorophore such that the antibody binds to a specific target molecule — which, in the case of the work on studying the proteomic signatures of synapses [40] going on in Smith’s lab, is generally a protein. This step is rate limited by the number of different fluorophores that one can simultaneously differentiate in the subsequent imaging step.

It is possible use solvents to break the bonds linking the target molecule and its complement antibody and wash away the dyes following one step of imaging to apply a different stain — that is to say, a stain involving new dyes with antibodies that target different molecules, proteins in the case of Smith’s work. However, you can only apply this “rinse and repeat” cycle so many times without significantly degrading the tissue sample. Currently Smith can image up to about 64 proteins before the tissue is compromised. How might you improve on this number? One possibility is to replace the fluorophore with a more complex molecule capable of encoding more bits. With seven fluorophores you can only encode the equivalent of three bits, but if you had a molecule of DNA with seven base pairs and could associate each target molecule/antibody with a unique sequence of seven nucleotides then you could distinguish between 4⁷ = 16,384 different molecules. Of course, you couldn’t simply read off the DNA sequence with EM, but you could slice each section into tiny cubes, sequence the DNA in each cube, assign each cube its coordinates in the tissue sample, and build a 3-D map annotated with the labels of the target molecules inferred from sequencing [3, 34, 39].

Sometimes help arrives from an unexpected quarter. Who would have thought we would still be discovering better ways to prepare neural tissue samples for subsequent analysis using increasingly sophisticated imaging techniques? The answer is, anyone really familiar with current work in histology and neuropathology. To an unsophisticated observer, it might seem that neuroanatomists of the late 19^th century such as Camillo Golgi (1843-1926), Ramon y Cajal (1852-1934) and Franz Nissl (1860-1919) had already made all the important breakthroughs. Many of their staining techniques are still in use today in one form or another. And so it came as somewhat of a surprise to non-specialists when it was announced in the April issue of Nature that a new preparation called CLARITY effectively renders a neural tissue sample — even an entire mouse brain — essentially transparent [14] allowing researchers to apply their full panoply of genomic, proteomic and connectomic tools.⁸

Pretty amazing technology! Of course, there’s no chance of recording any dynamics, but still it could be a real game changer for the field of connectomics. Recent work [2] from Janelia Farm also published in April on imaging transparent zebrafish does offer opportunities for observing the dynamics of neural activity. In this case, it is only the skin that is transparent and then only during the larval stage, but there are now also transgenic lines that retain their transparency as adults. Given its small size — 3 mm long, 2 mm thick, and 2.5 mm wide, once the skin is made transparent, two-photon microscopy can be used to image the entire brain of a living zebrafish.⁹

There are a lot of interesting technologies waiting in the wings involving both biological molecules and nanotechnology devices. The trick is combining them to provide an end-to-end technology that actually scales. In the near- to intermediate-term, it appears that so-called “imaging” technologies will be the first to provide significant improvements in scale. I say “so-called” because in many cases they don’t produce images per se, but rather they use energy radiated in the electromagnetic or acoustic spectrum to first illuminate the target tissue and then analyze the reflected energy to read off and localize cellular information.

The methods we’ve looked at so far rely on the visible range of the electromagnetic (EM) spectrum. Light in this range cannot penetrate deeply into tissue. At radio frequencies below 4 MHz the body is essentially transparent to the energy which makes it a candidate for transmitting data but not imaging. Light in the near-infrared range with a wavelength of about 800 nm to 2500 nm can penetrate tissue to several centimeters and is used in both imaging and stimulating neural tissue, and, as we’ll see in a few more slides, the electromagnetic spectrum is not the only alternative we have for non-invasively probing the brain.

The above graphic illustrates four different modalities for recording from the brain [36, 17]. The first three utilize different signals in the electromagnetic spectrum. The fourth utilizes molecular machinery to sense and record chemical and electrical behavior at the cellular level.

(a) The simplest recording device used by electrophysiologists is a metal needle insulated everywhere but at its tip, which is inserted into living tissue so as to rest in close proximity to a neuron and used to measure extracellular local field potentials to predict action potentials. The insulated portion of the needle is called the shank and the tip is called the electrode. There are many variations on this basic theme often involving multiple electrodes spaced along the shank that are used to record from several sites simultaneously. Current state-of-the-art technology supports on the order of 1000 electrodes, arranged linearly along a single shank [20], in a planar configuration usually referred to as a multi-electrode array [35] (MEA), and, more recently, in devices consisting of many multi-electrode shanks arranged in several rows like the bristles of a brush with the base of each shank anchored to a rectangular baseplate thereby allowing for the 3-D placement of thousands of electrodes [54]. The biggest disadvantage of this technology is that it is invasive and used in human subjects only when necessary, as in the case of guiding a surgeon in removing a brain tumor. Resolution: 10 micron resolution in z, 100 micron resolution in x and y, and millisecond temporal resolution. Improvements in x and y resolution are difficult due to tissue displacement and cellular damage resulting from device insertion.

(b) The second option is to use light of an appropriate frequency to illuminate the target tissue penetrating to such depth as the signal allows, and then analyze the reflected signal to infer properties of the cells and their associated molecules and behavior. The method of one- and two-photon microscopy using near infrared laser light is one of the more powerful technologies associated with this approach. In this method, molecules of interest are tagged with fluorescent dyes and then illuminated by laser light at a frequency such that one or two photons, depending on which variant of the technology is being employed, are sufficient to excite an electron in the dye molecule such that, when it relaxes from its excited energy state back to its ground state in the span of a few femtoseconds, it emits a photon of a shorter wavelength than the excitation photon(s). In the case of two-photon microscopy, the laser beam is focused on a small spot and scanned line-by-line over the sample tissue [26]. As a result, only dye molecules within the focal spot are excited. While a very useful technology, even with a near-infrared (NIR) laser operating betwen 800 nm to 2500 nm, penetration is limited to a few centimeters at best and thus the technology has limited potential for recording from large-brained, awake, behaving animals. However, there is great promise for studies involving mice [24]. Resolution: ~0.5 mm² field of view, and ~2.5 micron lateral (x and y) resolution.

(c) Nuclear magnetic resonance imaging allows us to observe and manipulate the magnetic properties of certain atomic nuclei. An MRI machine uses a powerful magnet to induce a magnetic field that serves to align the spins of magnetic nuclei. A resonant radio-frequency (RF) pulse is used to produce a second magnetic field perpendicular to the first which perturbs the alignment causing the nuclei to precess much like a top spinning in a gravitational field. The frequency of precession depends on the properties of the nuclei and the strength of the magnetic field. We measure the RF signals emitted by the spinning nuclei using an RF receiver and a special coil serving as an RF antenna called a probe located inside the MRI machine and used to predict the precession frequency and thereby infer properties of molecules in a tissue sample. Not all atomic nuclei have the right magnetic properties thus limiting what sort of molecules one can detect — hydrogen does, allowing us to study the movement of water molecules. In addition, the signals we are trying to measure are very weak and hence the challenge is to build a device with high-enough signal-to-noise ratio in a given localized region of the brain, and with a high-enough spatial resolution to measure the properties of molecules at the cellular level. State-of-the-art spatial resolution using experimental coils — not yet available in production machines — is on the order of 100 × 100 × 500 μm³ [27]. Temporal resolution depends on what you’re trying to measure and how much you have to sample from a region within a given recording volume or voxel to make measurements of the required accuracy [28, 6]. Resolution: MRI — 1 s temporal, 5 mm spatial; MEG — 1 ms temporal, 1 cm spatial; EEG — 1 ms temporal, 5 cm spatial.

(d) Biological organisms routinely generate, encode, transcode, transmit and decode electrical and chemical signals. The machinery for doing so is now being adapted to serve diverse technological purposes from sequencing DNA as in the case of third-generation sequencing technologies using nanopores [47] or optogenetics in which naturally occurring light-gated ion channels originally used to excite or inhibit neural activity [8, 51] are now being developed as voltage sensors capable of resolving action potentials [7]. Methods for encoding digital information in the form of sequences of nucleotides in strands of DNA have been around for some time [16] and now we are seeing new ideas about using molecular machinery to measure proxies for electrical activity and store the resulting information on DNA  [32, 53]. We’ll explore these and other related applications in more depth later in the talk.

Ultrasound pressure waves in the 2 to 18 MHz range are routinely used in diagnostic medical imaging and easily penetrate tissue to provide real-time images of the cardiovascular system. Ignoring temperature and barometric pressure, ultrasound travels about five times as fast in water (about 1500 m/s) than it does in air (about 300 m/s) and brain tissue behaves pretty much like water. Sound and pressure play an important role in a number of brain sensing technologies and many of them exploit the piezoelectric effect in one way or another. When materials such as ceramics, bone and DNA are mechanically stressed, they accumulate an electrical charge. Conversely, by applying an external electric field these materials, we can induce a change in their static dimensions. This inverse piezoelectric effect is used in the production of ultrasonic sound waves.

High-intensity focused ultrasound developed for medical applications employ phased-array ultrasonics in which an array of piezoelectric transducers is used to produce multiple pressure waves whose phase is adjusted by introducing delays in the electrical pulses that generate the pressure waves. By coordinating these delays, the focal point — point of highest pressure and thus highest temperature in the tissue — can be precisely controlled. The result is an instrument that can destroy a brain or breast tumor without cutting into the surrounding tissue. In the case of the brain, the cranium poses a challenge due to its variable thickness, but this can be overcome either by performing a craniotomy or by using a CT scan to construct a 3-D model of the skull and then generating a protocol based on this model that adjusts the delays to correct for aberrations in signal propagation due to the changes in thickness of this particular skull.

Sadek et al [43] employ the piezoelectric effect to design a multiplexer that makes it possible to build neural probes that have tens of thousands of sensors along their length. Existing state-of-the-art sensor are limited in the number of sensors they can support due to the requirement of etching a separate parallel trace to carry the signal for each sensor. The Sadek et al multiplexer works by transmitting multiple signals along a single channel using an idea inspired by the cochlea in which cochlear hair cells are implemented as nanoscale piezoelectric devices. Ultrasound is a versatile tool that be used for both stimulating and imaging the brain.

Wang et al [49] use focused ultrasound and a hybrid near-infrared-plus-ultrasound-imaging technology called photoacoustic spectroscopy to monitor focused-ultrasound-induced blood-brain barrier opening in a rat model in vivo. The very idea that focused-ultrasound (FUS) can induce changes in the blood-brain barrier (BBB) that would facilitate the passage of nanoparticles to deliver drugs is interesting in and of itself. One problem with this method is that the degree of BBB opening varies substantially and it is important to be able to monitor changes in the BBB in order to control dosage. Currently such monitoring is done with contrast-enhanced MRI, but the spatial and temporal resolution of this method is a limiting factor. The Wang et al monitoring procedure is complicated but I’ll try to provide a succinct high-level description.

First, the FUS method of induced BBB opening is applied — the rat undergoes a craniotomy to aid in beam focusing and power modulation, and then, prior to sonication, an ultrasound contrast agent (SonoVu [18] developed by Bracco Diagnostics) is intravenously injected to facilitate acoustic cavitation. The details regarding power and pulse length provide interesting insight into how to control for adverse effects in the test animals. Next AuNRs¹⁰ (gold nanorods) are injected into the jugular vein and accumulate at the BBB-opening-focus following sonication. Finally the area to be monitored is illuminated by a tunable laser¹¹ and then scanned by an ultrasonic transducer using a set of piezoelectric drive motors with a step size of 120μm in each of two directions. The authors claim that the “experimental results show that AuNR contrast-enhanced photoacoustic microscopy successfully reveals the spatial distribution and temporal responses of BBB disruption area in the rat brains.”

Imagine one type of reporting device distributed throughout the brain so there is one such device within a few microns of every neuron. Each such reporter is responsible for picking up the signals broadcast from recorders located on neurons in its immediate vicinity. These reporting devices might directly transmit information to receivers located outside the brain but we anticipate this approach running into power and transmission limitations. Instead, suppose that these local reporters participate in a heterogeneous network of devices allowing for different modes of transmission at different scales.

The local reporters would forward information to a second type of reporting device more sparsely distributed and responsible for relaying the information received from the local reporters to external receivers not limited by size or power requirements. These relay devices could number in the thousands instead of millions, be somewhat larger than the more numerous and densely distributed reporters, and be located inside of the dura but on the surface of the brain, within fissures or anchored on the membranes lining the capillaries supplying blood to the brain and the ventricles containing cerebrospinal fluid.

We anticipate and analyze one such scheme in [17]. In line with our discussion of the acoustic spectrum, the above figure illustrates a second approach in which the sub-dural reporters communicate with and supply power to another class of (implanted) reporters using ultrasonic energy. In the paper describing this work [44], the implanted reporters are referred to as neural dust and we’ll refer to an individual implanted reporter as a dust mote. Dust motes are also responsible for recording measurements of the local field potential (LFP) that the dust motes then transmit to the sub-dural reporters. The dust motes are approximately cube shaped, measuring 50 μm on a side, and consist of a piezoelectric device and some electronics for powering the mote and transmitting LFP measurements.

A 100 μm mote designed to operate at a frequency of 10 MHz, λ = 150 μm has ~1 dB attenuation at a 2 mm depth — the attenuation of ultrasound in neural tissue is ~0.5 dB/(cm MHz). A comparable electromagnetic solution¹² operating at 10 GHz, λ = 5 mm has ~20 dB attenuation at 2 mm. There are other complicating factors concerning the size of the inductors required to power EM devices and efficient antennae for signal transmission — displacement currents in tissue and scattering losses can increase attenuation to 40 dB in practical devices.

The small size of the implanted devices also complicates directly sensing the local field potential. The electrodes employed by electrophysiologists measure local field potentials at one or more locations with respect to a second, common electrode which acts as a ground and is located at some distance from the probe. In the case of LFP measurements made by dust motes, the distance between the two electrodes is constrained by the size of the device such that, the smaller this distance, the lower the signal-to-noise ratio — see Gold et al [25] for an analysis of extra- and intra-cellular LFP recording strategies for constraining compartmental models. The S/N problem can be ameliorated somewhat by adding a “tail” to each dust mote thereby locating the second electrode at some distance from the first one situated on the main body.

To get a better idea of the scale of the problems we’re considering, here are some numbers that quantitative neuroscientists keep in mind when doing back-of-the-envelope calculations. 100 billion of anything is a lot, but 100 billion sophisticated computing machines is staggering. White matter consists mostly of glial cells and myelinated¹³ axons covered with an insulating sheath that speeds transmission and ameliorates the effects of noise and the potential for crosstalk.

The number of neurons is perhaps less important than the number of active connections or synapses. Scott McNealy at SUN Microsystems was fond of saying “It’s the network stupid”, and his statement applies to computing in the brain as well as computing networks that characterize modern cloud computing architectures. There are something on the order of 1000 trillion synapses in a human brain and the molecular machinery operating at these connections is similarly complex.

These lists of quantitative facts just scratch the surface of what you’ll need for any particular analysis. The Harvard BioNumbers site provides an extensive searchable database of useful numbers. For instance, I recently wanted to know about nucleotide misincorporation rates and this query provided me with the information I was looking for.

Certainly we are not going to position a probe at every location within a few microns of a synapse in the human brain. We may, however, be able to develop nanoscale machines and distribute them throughout the brain so that every neuron can comfortably accommodate a few thousand of these machines positioned strategically in its active synapses. These machines would be designed to record the passage of proteins called neurotransmitters, which are the primary currency for exchanging information between neurons. Fortunately, given the state of the art in molecular biology, we don’t have to generate these machines de novo, but rather it seems plausible that we will be able to adapt existing cellular machinery from a variety of organisms to perform the basic sensing and communicating tasks required.

Every cell in the human body, and neurons in particular, contains a collection of molecular machines, some free floating and independent, and others anchored in cellular structures called organelles that serve as factories for the production and shaping of proteins, lipids and other macromolecules. In this graphic, the structure labeled (2) is the nucleus of the cell containing the DNA instructions to build the entire organism. (1) is called the nucleolus and is responsible for transcribing (ribosomal) RNA used to build ribosomes — shown as small dots one of which is labeled (3) — that are housed in the endoplasmic reticulum (5) a structure that serves to fold the newly minted but only partially-formed proteins into their final conformations. There are also organelles, the Golgi apparatus (6), that manufacture the many types of membranes that make up the structural members of the cell and a class of organelles called mitochondria (9) which are actually symbiotic organisms with their own separate DNA that set up housekeeping within the cells of eukaryotes many millions of years ago and made themselves indispensable.

All of these molecular machines are constantly at work performing the specialized functions of the particular cell type as well as performing a range housekeeping and routine repair chores. Everything that happens in the cell depends on the manufacture of proteins and amino acids using the cell’s DNA — its genetic code — as a blueprint. Indeed, the genetic code is being read and decoded at every moment in every cell in your body. You can’t move a muscle or think a thought without the production of scores of proteins and amino acids that work together to produce a dizzying array of behaviors manifest across an impressive range of temporal and spatial scales. Here’s an artistic rendering of a ribosome showing its two major molecular structures and a hint at its complicated shape and function.

Ribosomes are molecular machines constructed from ribosomal RNA and additional proteins that work in concert with transfer RNAs to pluck amino acids out of the fluid or cytoplasm that comprises much of the cell’s interior and consists of water and various dissolved molecules. These amino acids are formed into polypeptide chains that are subsequently folded to produce the specific three-dimensional shape or conformation that determines the protein’s function.

There are also molecular machines that serve to transport proteins from their place of manufacture, primarily organelles in a region called the soma in a neuron, to distant locations in the axons and dendrites that are involved in transferring information between neurons. Molecules called kinesins transport these protein products along pathways called microtubules.

They perform this essential service in the cell by utilizing molecules of stored energy to flex the kinesins, thus altering their conformation and enabling the kinesin molecules to essentially walk along the microtubules carrying their protein cargoes. In the Berkeley lecture I showed some clips from the “Inner Life Of A Cell” produced by BioVisions at Harvard. If you want to see just the clips I used, fast forward 3:20 into the video for the clip on microtubules, transport vesicles, kinesins (motor proteins), and 4:40 for the clip on ribosomes, ribonucleic acids, ribonucleotides and amino acids.

It is said that if you can imagine an operation that might be performed in a cell involving some manipulation of proteins or amino acids, then there is almost certainly some organism whose cells routinely perform that operation. Natural selection has had billions of years in which to explore the possibilities and seldom misses a trick. Molecular biologists are amassing a great catalog of such machines, many of which can be adapted to perform functions in cells other than those in which they are found naturally. This means that often as not if we need a molecular machine to perform a particular function we can order one from this catalog and adapt it to suit our purposes.

A good example of such adaptation comes from the latest generation of devices for sequencing DNA. The Human Genome Project more or less completed the sequencing of the human genome in 2004 after more than a decade of work and a cost of nearly 3 billion dollars. But this was just a single instance of the genome — actually it was a patchwork of pieces of DNA from several individuals, and, while we share a good deal of our individual genetic code, it is the differences between individuals that are likely to provide the clues in finding the causes and cures for many diseases. The race was on to drive down the cost and reduce the time required to days if not minutes.

The early gene sequencing technology required a good deal of machinery to automate what scientists had done on a smaller scale at their lab benches using diverse reagents and complex preparations. Some of the companies offering genome sequencing imagined building warehouses full of computers and robots programmed to carry out biochemical assays. It is worth noting that every cell in your body is sequencing your genome every minute of every day. How hard could it be? Several companies in an effort to scale sequencing are working to reduce the most time consuming part — reading off the sequence of nucleotides¹⁴ in a single strand of DNA, to a process that can be carried out on a silicon chip.

Oxford Nanopore has developed a technology that inserts a protein channel or pore in a silicon substrate. Single-strand DNA is threaded through this pore and a sensor reads out the voltage resulting from molecular changes as each nucleotide passes through the sensor.

The voltage is correlated with the type of nucleotide — one of adenosine, cytidine, guanosine or thymidine, abbreviated as A, C, G and T in the case of DNA, so that the sequence can easily be read off.

This approach scales so that you can populate a single silicon die with thousands of these pores arranged in a regular grid. It is also possible to sculpt a tiny hole in the silicon with a laser to achieve the essential properties of the protein required for sequencing. Aside from preparing and cloning the original sample of DNA which can be accomplished using PCR, this technology avoids most of the complicated chemical processes and expensive reagents that plagued earlier technologies¹⁵.

Another example of re-purposing existing molecular machinery comes to us from recombinant DNA technology associated most closely with DNA cloning. In one of the simpler methods applications of recombinant DNA technology, a target piece of DNA is assembled in a self-replicating package called a plasmid and introduced into an easy-to-grow bacterial host such as E. Coli. The host cells containing the plasmid are then grown in a nutrient medium using the host’s reproductive machinery to produce a large quantity of the modified organism.

Francis Crick — of Crick and Watson, double-helix fame — challenged neuroscientists to determine the connectivity pattern of a single neuron and all of its neighbors. This is a special case of a connectome which we mentioned earlier. To accomplish this feat, researchers applied another recombinant DNA tool to infect progenitor stem cells that can grow into neurons with a self-replicating virus called a retrovirus. A retrovirus is an RNA-based virus that replicates in a host cell using one of its own enzymes called reverse transcriptase to produce DNA from its RNA genome. The viral DNA is then incorporated directly into the host’s genome using another enzyme called integrase. The virus thereafter replicates as part of the host cell’s DNA.

HIV and rabies are particularly nasty instances of retroviruses that have been adapted for (synthetic) recombinant DNA purposes. These viruses are loaded with a DNA payload which they splice into the host’s DNA and thereby use the host’s sequencing and protein-transcribing machinery to reproduce the virus or carry out any other process that can be accomplished by the cellular machinery. In the case of addressing Crick’s challenge, the virus uses the microtubule-based transport system we described earlier to propagate copies of itself throughout the cell and infect adjacent cells connected through axonal processes. Fluorescent markers attached to the viral DNA are used to highlight the connectome and produce the images you see here.

Retroviruses and recombinant DNA technology also provide powerful tools for controlling individual neurons. Optogenetics is the name given to a collection of tools that use light to activate or silence neurons [8, 9]. The electrical properties of neurons and the action potentials that propagate electrical signals along axonal processes are governed by differences in ion concentrations across the cell membrane. Voltage- and ligand-gated ion channels — a ligand is a biomolecule involved in cellular signalling and neurotransmitters constitute the class we are most concerned with here — serve to alter ion concentrations by allowing specific ions — sodium, potassium and calcium — to flow through the membrane and reestablish the equilibrium in accord with the prevailing voltage and concentration gradients. Our cellular machinery is exquisitely tuned so that only a few ions are required to cause substantial changes in the voltage difference across the axonal membrane.

Some species of algae have light-gated ion channels that can be introduced into neurons using recombinant DNA technology. If you shine a light of the proper frequency on these channels they open, altering the potential across the cell membrane and, depending on the voltage drop, initiating an action potential to propagate along the the axon. There is another technique in the optogenetics toolkit that relies on a different wavelength of light to silence a neuron, thus preventing it from initiating a spike, and thereby enabling the researcher to both turn on and turn off individual neurons. This approach to controlling neurons is extremely useful in establishing cause-and-effect relationships, but it still requires additional instrumentation, such as conventional voltage-sensing probes to measure electrical activity in neurons.

The use of tame variants of naturally occurring retroviruses is a powerful tool for co-opting cells to perform tasks required for the production and delivery of drugs and gene therapies. The field of synthetic biology focuses on engineering biological circuits¹⁶ consisting of standard components that can be ordered from a catalog and depended upon to reliably produce a specified behavior. Ultimately, the goal is to build molecular machines using such components so that all the operations required for a given application are performed entirely within our living cells [15].

Researchers in the field of synthetic biology are now able to build circuits of biomolecules that perform computations such as logical operations that are not readily accessible in nature. DNA polymerase is the workhorse for building such circuits. This enzyme, which is found in all cells including bacteria and viruses, synthesizes DNA molecules from their nucleotide building blocks. Circuits using such naturally occurring enzymes are remarkable efficient, but they are quite slow. One of our collaborators has designed a three-bit demodulator circuit that requires over 300 polymerase reactions. Silicon logic is fast but it is also incredibly inefficient by comparison.

In the case of the nanopore gene sequencing technology, a sample is extracted from the target organism and the sequencing is performed externally. In the case of tracing the connectome using a rabies virus, the experiments are performed in vitro using cells grown in a culture. How might we operate directly on cells in vivo, that is to say in live animals without requiring surgery, inserting probes or sacrificing the animal to extract information for analysis.

One possible answer is to use recombinant DNA technology to re-program cells to perform new operations while not interfering with their normal function. Where in the case of employing a retrovirus to trace the connectome we used the existing inter-cellular transport mechanism, in the case of delivering a viral vector to a large population of target cells — neurons in our case — we might exploit the system of arteries and capillaries that deliver nutrients and oxygen to every cell in the human body.

In the particular case of the brain, delivery is complicated by the intervention of the blood-brain barrier which consists of a membrane called the endothelium that surrounds each capillary and through which every molecule entering the brain must pass. Once inside the brain glial cells called astrocytes assist in the exchange of oxygen and glucose and the production of enzymes and neurotransmitters essential for neurons to perform their duties.

The blood brain barrier protects the brain from toxins and pathogens that might disrupt the neural machinery controlling vital processes throughout the body. Unfortunately for those affected by viral-borne brain diseases, nature has figured out how to bypass the barrier and the HIV and rabies retroviruses mentioned earlier are examples of such pathogens. The silver lining is that we are figuring out how to use these same viral vectors to repair cell damage and deliver drug payloads selectively to targets throughout the body and the brain in particular. We’re also figuring out ways to foil natural viruses so they can’t cross the blood-brain barrier.

If we can use the arteries and capillaries to distribute and deliver molecular machines, then we might also use the lymph system and the vessels that return blood to the lungs and heart and waste products to the kidneys as a means of conveying information to locations external to the central nervous system where it might be more easily processed, say, using an artificial filtering process akin to dialysis¹⁷. This would provide an expedient for reading out neural-state information in lieu of more complicated nanotechnology solutions employing tiny radio transmitters that have been suggested in the literature.

Now we have a means of providing input and extracting output from the brain¹⁹. Granted that the input modality we’ve been exploring requires we infect each cell with virus and modify its DNA, and that would likely not serve as the input side of a real-time computer interface. On the output side however, we might have a shot at being able to observe a behaving brain at an unprecedented scale and level of detail. Assuming that our virally-delivered molecular machines don’t interfere with the normal operation of the cells, we could in principle develop technology for reading off states of the brain that would not harm the host and could operate indefinitely²⁰. What sort of information might we want to collect and how would we go about doing so?

Traditionally the focus has been on recording spike trains in the form of changes in the membrane potential of individual neurons. However, the signaling pathways in the brain are subtle and multitude; they include electrical pathways²² in the form of action potentials and voltage-gated ion channels, genetic pathways in the form of DNA translated into RNA and proteins expressed and transported within the cell, and chemical pathways in form of neurotransmitters which are emitted into the synaptic cleft separating an axon and a dendrite and serve to open ligand-gated ion channels on the dendrite.

Here we see a schematic synapse showing the neurotransmitters packaged in vesicles (2) in the pre-synaptic neuron A, ligand-gated ion channels (5) in the post-synaptic neuron B, and a mitochondrial organelle (1) that provides energy in the form of ATP. Other components include voltage-gated calcium²³ channels (6) that are activated by action potentials and cause the vesicles to merge with the cell membrane and neurotranmitters to flood into the synaptic cleft and additional machinery responsible for scavenging neurotransmitters in a process called reuptake.

Must we record the state of all these components in order to obtain a complete picture? Perhaps, but it may be that the proteomic history — the record of specific proteins expressed and transported across synapses — is sufficient to infer most of what is going on informationally and computationally within the brain. In any case, we are going to assume so for the remainder of this discussion, and make the additional simplifying assumption that the production and transfer of neurotransmitters provide enough information.

Our grand goal is to collect enough data to infer not only the structure and circuitry of the brain — what we have been calling the connectome, but the function of smaller, anatomically-localized neural circuits, larger super-complexes of neurons that implement functional areas such as the visual cortex, and the recurrent pathways linking these functional areas to support high-level cognition. We hope to abstract the behavior of these diverse neural circuits and build models to test our understanding. Our progress so far suggests that we will have to record simultaneously from large collections of neurons to make additional progress on these challenging problems.

Here’s a very rough sketch for how we might record the proteomic history of a behaving brain. First off we need to be able to identify what neurotransmitter is being conveyed, which neuron is transmitting the information and which neuron is receiving it. Essentially we need a unique identifier for each class of neurotransmitter and each individual neuron. A recent paper in Nature described a scalable method for generating self-assembling barcodes using DNA origami as a substrate and fluorescent tags to encode the digital information [34]. Something like this method might suffice to encode the unique identifiers that we require.

Next we need to associate neurotransmitters with their identifying barcodes and convey these barcodes along with the neurotransmitters making sure that they find their way into the receiving neuron where they can be assembled into packets that describe each event as a triple of three barcodes encoding the transmitting neuron, the receiving neuron and the class of neurotransmitter conveyed. Once assembled these packets would be flushed into the cerebrospinal fluid to be subsequently eliminated from the brain via the lymph and blood circulation system.

Fleshing this out would require a great deal of speculation, and so to reduce the hand waving to a tolerable level let’s consider one more-or-less concrete proposal suggested in a paper [52] out of Tony Zador’s group at Cold Spring Harbor Laboratory. The authors of this paper propose the idea of sequencing the connectome in their paper of the same title. They break down the problem into three components: (a) label each neuron with a unique DNA sequence or barcode, (b) propagate the barcodes from each source neuron to each synaptically-adjacent sink neuron — this results in each neuron collecting a “bag of barcodes”, and (c) for each neuron to combine its barcodes in source-sink pairs for subsequent high-throughput sequencing.

It would seem the authors have in mind sacrificing the animal in the final step, but it may be possible to pass source-sink pairs through the cell membrane and into the lymph-blood system for external harvest as proposed earlier. They suggest that propagation might be accomplished using a trans-synaptic virus such as rabies [41]. Additional techniques would be required to map barcodes to brain areas. The authors claim to be developing an approach based on PhiC31 integrase for joining barcodes and PRV amplicons [42] for trans-synapatic barcode propagation. While quite ambitious with many complicated steps yet to be filled in, a number of neuroscientists, myself included, believe that some variant of this idea could be accomplished in the relatively near-term future, and there are plans afoot [4] to take on even more ambitious goals²⁴.

The Zador et al proposal offers an interesting scalable approach based on a existing biological mechanisms for reading off the connectome using DNA sequencing technology. In a similar spirit, Zamft et al [53], building on some earlier work by Church and Shendure [16] and Kording [32], propose a method that may offer our best near-term solution to recording cellular state information at scale. Cellular transcoding processes such as DNA replication and messenger RNA transcription make occasional errors in the process of incorporating nucleotides into their respective products. Zamft et al propose a method for building a molecular recording device that exploits this otherwise undesirable property. Their method depends on harnessing an enzyme called DNA polymerase that plays a central role in DNA replication.

This HHMI animation of DNA replication provides a dramatic illustration of the potential power of harnessing DNA polymerase in this fashion. The process involves an enzyme that unwinds the DNA, and other enzymes that copy the two resulting strands. Both the strands of the DNA double helix act as templates for the new DNA strands. Incoming DNA is unraveled in segments by the enzyme helicase, resulting in one strand starting with a free 3’ hydroxyl group and ending with a 5’ hydroxyl group and a second strand with the two hydroxyl groups reversed. These two strands are replicated by DNA polymerase but in different directions since DNA polymerase only works on strands that begin with a 3’ hydroxyl group. The result is this remarkable example of cellular choreography.

The researchers working on the method described in [53] rely on the property that the rate of misincorporation is dependent on the concentration of positive ions or cations, specifically the concentration of Ca⁺² ions. They show that by modulating cation concentrations one can influence the misincorporation rate on a reference template in a reliable manner so that information can be encoded in the product of DNA polymerase and then subsequently recovered by sequencing and comparing with the reference template.

More often than not when I mention the idea of a technology that would require introducing millions of nanoscale machines into your brain using a genetically-engineered virus I get a fair bit of skepticism and, so far, no one has volunteered to beta test the technology.

I’ll leave you with this scene from The President’s Analyst in which James Coburn plays a psychiatrist confronted by the president of the phone company asking the Coburn character to use his influence with the president to help pass legislation requiring a new nanoscale phone technology to be implanted in the brain of every newborn. (You can check out the clip in this YouTube video.)

P.S. I could literally go on for hours and have in some more intimate classroom situations. I’ve created this annotated transcription of the talk with lots of footnotes providing additional detail and relevant papers for your further reading. I believe wanting to tell everyone about what you’ve discovered is a wonderful child-like characteristic that is also incredibly valuable to both the individual and society. It is a form of public thinking that is under-appreciated and often discouraged in precocious children. I encourage it in everyone I meet and cherish it in my students and colleagues. I thank you for your patience indulging me in my habit.

References

[1]	Leonard M. Adleman. Molecular computation of solutions to combinatorial problems. Science, 266(5187):1021–1024, 1994.
[2]	Misha B Ahrens, Michael B Orger, Drew N Robson, Jennifer M Li, and Philipp J Keller. Whole-brain functional imaging at cellular resolution using light-sheet microscopy. Nature methods, 10(5):413–420, 2013.
[3]	A. Paul Alivisatos, Anne M. Andrews, Edward S. Boyden, Miyoung Chun, George M. Church, Karl Deisseroth, John P. Donoghue, Scott E. Fraser, Jennifer Lippincott-Schwartz, Loren L. Looger, Sotiris Masmanidis, Paul L. McEuen, Arto V. Nurmikko, Hongkun Park, Darcy S. Peterka, Clay Reid, Michael L. Roukes, Axel Scherer, Mark Schnitzer, Terrence J. Sejnowski, Kenneth L. Shepard, Doris Tsao, Gina Turrigiano, Paul S. Weiss, Chris Xu, Rafael Yuste, and Xiaowei Zhuang. Nanotools for neuroscience and brain activity mapping. ACS Nano, 7(3):1850–1866, 2013.
[4]	A. Paul Alivisatos, Miyoung Chun, George M. Church, Ralph J. Greenspan, Michael L. Roukes, , and Rafael Yuste. The brain activity map project and the challenge of functional connectomics. Neuron, 74, 2012.
[5]	Björn Andres, Ullrich Köthe, Thorben Kröger, Moritz Helmstaedter, Kevin L. Briggman, Winfried Denk, and Fred A. Hamprecht. 3-D segmentation of SBFSEM images of neuropil by a graphical model over supervoxel boundaries. Medical Image Analysis, 16(4):796–805, 2012.
[6]	T. Atanasijevic, M. Shusteff, P. Fam, and A. Jasanoff. Calcium-sensitive MRI contrast agents based on superparamagnetic iron oxide nanoparticles and calmodulin. Proceedings of the National Academy of Science, 103(40):14707–14712, 2006.
[7]	L. Barnett, J. Platisa, M. Popovic, V.A. Pieribone, and T. Hughes. A fluorescent, optogenetic voltage sensor capable of resolving action potentials. PloS ONE, 7:e43454, 2012.
[8]	Jacob G. Bernstein and Edward S. Boyden. Optogenetic tools for analyzing the neural circuits of behavior. Trends in Cognitive Sciences, 15:592–600, 2011.
[9]	Edward S Boyden, Feng Zhang, Ernst Bamberg, Georg Nagel, and Karl Deisseroth. Millisecond-timescale, genetically targeted optical control of neural activity. Nature Neuroscience, 8:1263–1268, 2005.
[10]	Stephen F. Bush. Nanoscale Communication Networks. Nanoscale Science and Engineering. Artech House, 2010.
[11]	Stephen F. Bush. Toward in vivo nanoscale communication networks: utilizing an active network architecture. Frontiers in Computer Science China, 5(3):316–326, 2011.
[12]	Quan Chen, S.L. Ho, and W.N. Fu. Numerical investigation of magnetic resonant coupling technique in inter-chip communication via electromagnetics-TCAD coupled simulation. IEEE Transactions on Magnetics, 48(11):4253–4256, 2012.
[13]	Hak Soo Choi, Wenhao Liu, Preeti Misra, Eiichi Tanaka, John P. Zimmer, Binil Itty Ipe, Moungi G. Bawendi, and John V. Frangioni. Renal clearance of nanoparticles. Nature Biotechnology, 25(10):1165–1170, 2007.
[14]	Kwanghun Chung, Jenelle Wallace, Sung-Yon Kim, Sandhiya Kalyanasundaram, Aaron S. Andalman, Thomas J. Davidson, Julie J. Mirzabekov, Kelly A. Zalocusky, Aleksandra K. Denisin Joanna Mattis, Sally Pak, Hannah Bernstein, Charu Ramakrishnan, Logan Grosenick, Viviana Gradinaru, and Karl Deisseroth. Structural and molecular interrogation of intact biological systems. Nature, 10, 2013.
[15]	George Church and Ed Regis. Regenesis: How Synthetic Biology Will Reinvent Nature and Ourselves. Basic Books, 2012.
[16]	George Church and Jay Shendure. Nucleic acid memory device. US Patent 20030228611, 2003.
[17]	Thomas Dean, Biafra Ahanonu, Mainak Chowdhury, Anjali Datta, Andre Esteva, Daniel Eth, Nobie Redmon, Oleg Rumyantsev, and Ysis Tarter. On the technology prospects and investment opportunities for scalable neuroscience. Technical Report 41324, Google, 2013.
[18]	P.A Dijkmans, L.J.M Juffermans, R.J.P Musters, A van Wamel, F.J ten Cate, W van Gilst, C.A Visser, N de Jong, and O Kamp. Microbubbles and ultrasound: from diagnosis to therapy. European Journal of Echocardiography, 5(4):245–246, 2004.
[19]	K. Eric Drexler. Nanosystems: Molecular Machinery, Manufacturing, and Computation. Wiley, New York, 1992.
[20]	J. Du, T.J. Blanche, R.R. Harrison, H.A. Lester, and S.C. Masmanidis. Multiplexed, high density electrophysiology with nanofabricated neural probes. PLoS One, 6(10):e26204, 2011.
[21]	Michael D. Fayer. Elements of Quantum Mechanics. Oxford University Press, Oxford, UK, 2001.
[22]	Michael D. Fayer. Absolutely Small: How Quantum Theory Explains our Everyday World. AMACOM, New York, NY, 2010.
[23]	Evan H. Feinberg, Miri K. VanHoven, Andres Bendesky, George Wang, Richard D. Fetter, Kang Shen, and Cornelia I. Bargmann. GFP reconstitution across synaptic partners (GRASP) defines cell contacts and synapses in living nervous systems. Neuron, 57(3):353–363, 2008.
[24]	K.K. Ghosh, L.D. Burns, E.D. Cocker, A. Nimmerjahn, Y. Ziv, A.E. Gamal, and M.J. Schnitzer. Miniaturized integration of a fluorescence microscope. Nature Methods, 8(10):871–8, 2011.
[25]	C. Gold, D.A. Henze, and C. Koch. Using extracellular action potential recordings to constrain compartmental models. Journal Compututational Neuroscience, 23(1):39–58, 2007.
[26]	F. Helmchen and W. Denk. Deep tissue two-photon microscopy. Nature Methods, 2(12):932–940, 2005.
[27]	M. C. Hsieh, L. W. Kuo, E. L. Wu, and J. H. Chen. Localized high-resolution MR imaging of rat brain architecture using micro-fabricated receive-only RF coil. In IEEE Engineering Medical Biology Society, pages 6074–6077, 2007.
[28]	V. Hsieh and A. Jasanoff. Bioengineered probes for molecular magnetic resonance imaging in the nervous system. ACS Chemical Neuroscience, 3:593–602, 2012.
[29]	Viren Jain, H. Sebastian Seung, and Srinivas C. Turag. Machines that learn to segment images: a crucial technology for connectomics. Current Opinion in Neurobiology, 20(5):1–14, 2010.
[30]	Richard Anthony Lewis Jones. Soft machines: Nanotechnology And life. Oxford University Press, Oxford, UK, 2008.
[31]	Suhasa B. Kodandaramaiah, Giovanni T. Franzesi, Brian Y. Chow, Edward S. Boyden, and Craig R. Forest. Automated whole-cell patch-clamp electrophysiology of neurons in vivo. Nature Methods, 9(6):585–587, 2012.
[32]	Konrad Kording. Of toasters and molecular ticker tapes. PLoS Computational Biology, 7(12):e1002291, 2011.
[33]	Mark S. Leeson. Communicating at the nanoscale. In Proceedings of the 5th WSEAS International Conference on Communications and information technology, pages 288–293. World Scientific and Engineering Academy and Society (WSEAS), 2011.
[34]	Chenxiang Lin, Ralf Jungmann, Andrew M. Leifer, Chao Li, Daniel Levner, George M. Church, William M. Shih, and Peng Yin. Submicrometre geometrically encoded fluorescent barcodes self-assembled from DNA. Nature Chemistry, 4:832–839, 2012.
[35]	A. M. Litke, N. Bezayiff, E. J. Chichilnisky, W. Cunningham, W. Dabrowski, A. A. Grillo, M. Grivich, P. Grybos, P. Hottowy, S. Kachiguine, R. S. Kalmar, K. Mathieson, D. Petrusca, M. Rahman, and A. Sher. What does the eye tell the brain?: Development of a system for the large-scale recording of retinal output activity. IEEE Transactions on Nuclear Science, 51:1434–1440, 2004.
[36]	Adam H. Marblestone, Bradley M. Zamft, Yael G. Maguire, Mikhail G. Shapiro, Thaddeus R. Cybulski, Joshua I. Glaser, Ben Stranges, Reza Kalhor, Elad Alon David A. Dalrymple, Dongjin Seo, Michel M. Maharbiz, Jose Carmena, Jan Rabaey, Edward S. Boyden, George M. Church, and Konrad P. Kording. Physical principles for scalable neural recording. ArXiv preprint cs.CV/1306.5709, 2013.
[37]	Kristina D. Micheva and Stephen J. Smith. Array tomography: A new tool for imaging the molecular architecture and ultrastructure of neural circuits. Neuron, 55(1):25–36, 2007.
[38]	Shawn Mikula, Jonas Binding, and Winfried Denk. Staining and embedding the whole mouse brain for electron microscopy. Nature Methods, 9:1198–1201, 2012.
[39]	R.D. Mitra, J. Shendure, J. Olejnik, E-K. Olejnik, and Church G.M. Fluorescent in situ sequencing on polymerase colonies. Analytical Biochemistry, 320(1):55–65, 2003.
[40]	Nancy A. O’Rourke, Nicholas C. Weiler, Kristina D. Micheva, and Stephen J. Smith. Deep molecular diversity of mammalian synapses: Why it matters and how to measure it. Nature Review Neuroscience, 10:365–379, 2012.
[41]	Emin Oztas. Neuronal tracing. Neuroanatomy, 2:2–5, 2003.
[42]	Jesús Prieto, Jesús Solera, and Enrique Tabarés. Development of new expression vector based on pseudorabies virus amplicons: application to human insulin expression. Virus Research, 89(1):123–129, 2002.
[43]	A.S. Sadek, R.B. Karabalin, J. Du, M.L. Roukes, C. Koch, and S.C. Masmanidis. Wiring nanoscale biosensors with piezoelectric nanomechanical resonators. Nano Letters, 10:1769–1773, 2010.
[44]	Dongjin Seo, Jose M. Carmena, Jan M. Rabaey, Elad Alon, , and Michel M. Maharbiz. Neural dust: An ultrasonic, low power solution for chronic brain-machine interfaces. ArXiv preprint cs.CV/1307.2196, 2013.
[45]	H. Sebastian Seung. Reading the book of memory: Sparse sampling versus dense mapping of connectomes. Neuron, 62(1):17–29, 2009.
[46]	Sebastian Seung. Connectome: How the Brain’s Wiring Makes Us Who We Are. Houghton Mifflin Harcourt, Boston, 2012.
[47]	Gregory Timp, Utkur Mirsaidov, Winston Timp, Jiwook Shim, Deqiang Wang, Valentin Dimitrov, Jan Scrimgeour, Chunchen Lin, Jeffrey Comer, AnthonyH. Ho, Xueqing Zou, Aleksei Aksimentiev, and Klaus Schulten. Third generation DNA sequencing with a nanopore. In Samir M. Iqbal and Rashid Bashir, editors, Nanopores, pages 287–311. Springer US, 2011.
[48]	C. Lee Ventola. The nanomedicine revolution, Part 1: Emerging concepts. Pharmacy and Therapeutics, 37(9):512–525, 2012.
[49]	Po-Hsun Wang, Hao-Li Liu, Po-Hung Hsu, Chia-Yu Lin, Churng-Ren Chris Wang, Pin-Yuan Chen, Kuo-Chen Wei, Tzu-Chen Yen, and Meng-Lin Li. Gold-nanorod contrast-enhanced photoacoustic micro-imaging of focused-ultrasound induced blood-brain-barrier opening in a rat model. Journal of Biomedical Optics, 17:061222, 2012.
[50]	R.M. White, A. Sessa, C. Burke, T. Bowman, J. LeBlanc, C. Ceol, C. Bourque, W. Dovey, M. Goessling, C.E. Burns, and L.I. Zon. Transparent adult zebrafish as a tool for in vivo transplantation analysis. Cell Stem Cell, 2:183–189, 2008.
[51]	O. Yizhar, L.E. Fenno, T.J. Davidson, M. Mogri, and K. Deisseroth. Optogenetics in neural systems. Neuron, 71(1):9–34, 2011.
[52]	Anthony M. Zador, Joshua Dubnau, Hassana K. Oyibo, Huiqing Zhan, Gang Cao, and Ian D. Peikon. Sequencing the connectome. PLoS Biology, 10(10):e1001411, 2012.
[53]	Bradley Michael Zamft, Adam H. Marblestone, Konrad Kording, Daniel Schmidt, Daniel Martin-Alarcon, Keith Tyo, Edward S. Boyden, and George Church. Measuring cation dependent DNA polymerase fidelity landscapes by deep sequencing. PLoS ONE, 7(8):e43876, 2012.
[54]	A. N. Zorzos, J. Scholvin, E. S. Boyden, and C. G. Fonstad. Three-dimensional multiwaveguide probe array for light delivery to distributed brain circuits. Optics Letters, 23:4841–4843, 2012.