Designing Neural Networks in Culture

Abstract

Technology has advanced to where it is possible to design and grow—with predefined geometry and surprisingly good fidelity—living networks of neurons in culture dishes. Here we overview the elements of design, emphasizing the lithographic techniques that alter the cell culture surface which in turn influences the attachment and growth of the neural networks. Advanced capability in this area makes it possible to design networks of desired complexity. Other issues addressed include the influence of glial cells and media on activity and the potential for extending the designs into three dimensions. Investigators are advancing the art and science of analyzing and controlling through stimulation the function of the neural networks, including the ability to take advantage of their geometric form in order to influence functional properties.

I. INTRODUCTION

Building a brain on a chip has caught the imagination of a growing number of researchers. The origins of the idea date to pioneering work of Thomas [1], Pine [2], and Gross [3] who showed 30 years ago the feasibility of recording from cultured neurons, myocytes and isolated ganglia with planar electrode arrays. A related technique, the recording and stimulation of activity from brain slices, has progressed from its initial demonstrations [4]–[6] to its current state where the use of multielectrode arrays (MEAs) is so widespread as to support multiple commercial vendors. Common to these efforts is the idea that stimulation and recording of these networks with large numbers of electrodes would lead to better understanding of the basic neuroscience of learning and memory, neural coding, and properties of signal propagation in neural networks. The technology also provides unique approaches to the understanding of disease states such as epilepsy and stroke, and has potential for screening for neuroactivity of drugs in development [7]. That the brain shows both strong structure as well as local randomness has enticed a number of investigators to pursue the means to grow neurons in patterns so as to influence their functional behavior.

The thesis of this paper is that technology has developed to the point where it is beginning to be appropriate to talk seriously about designing a brain on a chip, not just investigating the properties of neural cell culture or brain slices. Most notably, the geometric pattern of growth can be controlled in dramatic fashion. However, there are a number of other design choices that are increasingly selectable by the neuroengineer. These start with the nature of the tissue (dissociated primary neurons vs. cell line vs. brain slice vs. ganglion) and its identity, and include the density of the cells, the composition of the media, the duration of the experiment, and the stimulation and recording protocols. It is even possible to construct three-dimensional culture systems, including recording and stimulation devices. This paper discusses these issues, with an emphasis on the lithographic techniques and the resulting neuron growth patterns.

Although the work described below is dominantly from our laboratories, there are researchers around the world involved in similar or complementary studies. The reader is referred to the MEA Conference Proceedings for a much more thorough list [8]. A preliminary version of this paper appeared as a conference abstract [9].

II. CELLULAR LITHOGRAPHY

Although patterning substrates to control cellular growth in culture has long roots [10]–[12] the field accelerated after the publication by Kleinfeld of the use of photoresist technology to pattern hydrophobic and hydrophilic materials to control neuronal cell attachment [13], as well as the introduction of UV photoablation [14]. The introduction of variety of techniques, including UV techniques [15], [16], photoresist patterning [17]–[19], microcontact printing [20]–[25], microfluidic deposition [26], [27], and micro-machined surfaces [28]–[30]. Materials patterned include hydrophobic alkyl- and hydrophilic amino-silanes on glass (and their thiol equivalents on gold), protein resistant polyethylene glycol, proteins, biological macromolecules, and critical peptide sequences. Linkers, including epoxy-silane [31], enhance the effectiveness of the surface chemistry. Substrates have included insulators glass, silicon, and various plastic polymers.

We have begun to understand the principles that control in vitro cell patterning. Initial observations are that hydrophilic materials are more conducive to cell attachment and growth (“cytophilic”), than “cytophobic” hydrophobic materials; positively charged materials are preferred to negatively charged molecules [13], [14]; growth is correlated with amine group density [32]. More complex materials, such as laminin and fibronectin, are known to control attachment. We have found a protein resistant material (polyethylene glycol) to be very successful at restricting neurite outgrowth [33], [34]. Some materials enhance axonal outgrowth, including mixtures of laminin and polylysine [35], [36] and the laminin derived peptide sequence P20 [37]. Topology is also important [28], [29], [38], [39].

Fig. 1 illustrates the patterning techniques and their use with electrode arrays. Most of the materials are compatible with multiple patterning techniques, including microcontact printing, photoresist patterning, laser ablation, microfluidic deposition, and microchannel deposition. However, to our knowledge only printing permits multiple different biomolecular cues to be applied to the surface. The surfaces may be any of the popular electrode array metals (gold, platinum, indium tin oxide, titanium nitride) and insulators (silicon nitride, silicon dioxide, glass, polyimide, PDMS). The materials deposited may be permissive to cell growth, including polylysine, laminin, or various aminosilanes, or they may be nonpermissive, including polyethylene glycol, albumin, and chondroitin sulfate. Of great interest, but still only recently exploited is the use of neural growth factors such as BDNF to provide precision guidance of neurons [40]. The biomolecules may be covalently linked through a variety of surface chemistry linkers, or they may be physisorbed onto the surface. Taken collectively, the experimenter has a very rich toolbox from which to choose.

Fig. 1.

Micropatterning technologies commonly used for cellular lithography. (a) Laser ablation: a permissive coating is ablated through a mask. (b) Photoresist processing: here a permissive material is protected in desired areas by photoresist; unprotected areas are chemically etched and then reacted with a non-permissive coating. (c) Micro-contact printing: the surface is prepared with a chemical linker, the stamp is inked with the permissive material and the surface stamped, perhaps with multiple inks, and perhaps reacted with a nonpermissive material in the final step. (d) Electrode arrays with metal conductors have their insulating surfaces modified to lead neurons to grow over the top of electrodes.

Confinement may be with physical channels or posts [41], [42] or wells [43] that restrict the movement of cell bodies while permitting axonal extension. Tunnels under PDMS [44] and agarose [45] have worked to provide pathways for networked neurons. As discussed below, the recent development of microtunnel/chamber structures show promise for isolating pure populations of neurons while permitting axonal communication [46], [47].

III. GEOMETRIC DESIGN CHOICES

The availability of working lithographic technology implies that the engineer must design the network. We are gaining experience in the ability to control growth patterns. Fig. 2 shows images of networks grown in our laboratories. These indicate that the choices include line width, node size, and geometric pattern. The range is from an approximation not too far from connecting individual neurons [Fig. 2(a), (b)] to using “bundles” of varying width [Fig. 2(c), (d)] to creating neuropil-like structure with connecting lines [Fig. 2(e)]. The crossing pattern [Fig. 2(f)] has dimensions comparable to Fig. 2(d).

Fig. 2.

Geometric patterns of rat hippocampal neuron growth in culture. (a) Network of individual neurons (laser patterned, Reproduced by permission of Wiley & Sons [14]. (b) Thin line network on 3 μm lines at 14 days in culture. (c) 10 μm wide line network at 55 days in culture. (d) 30 μm lines and 80 μm square nodes at 21 days in culture. (e) Neuropil structure separated by 500 μm with 3 μm wide lines. (f) Cross pattern of 80 μm nodes and 30 μm lines, stained for neurons (green), astroglia (red), and nuclei (blue) (figure was used with permission of the first author for the cover on the Journal of Neural Engineering). In (c), (d) epoxy-silane served as linker for stamped poly-D-lysine and as cytophobic background [31].

Networks of single or limited numbers of neurons may be too fragile for practical experimental use. Rutten’s group has studied the dependence of neural activity on size of clusters of neurons, emphasizing that a critical population size of neurons is needed to support both activity and survivability. They report delayed onset of activity and gradual breakdown of pattern (e.g., one month) [48], [49]. Anecdotally we have found that lines that are too thin provide too little substrate for robust attachment and survival of neurons. In one study, we observed migration of neural cell bodies to the larger nodes and measured compliance to pattern [14]. In another we found that isolated lines thinner than 25 μm wide resulted in clumping of neurons and insufficient electrical spontaneous or evoked activity to permit usage [50], [51]. A third study reported that when local cell densities on patterns exceeded 250 cells/mm², there was substantial activity [52]. With these results one of our design choices is to utilize 80 μm square nodes and 25 μm wide lines, as shown in Fig. 2(d), (f). However, an investigation with 4 μm wide lines and 10 μm diameter nodes, in conjunction with patch clamping, was quite successful for identifying circuit connections of individual cells [53]. For reference, the networks in Fig. 2(c), (d), (f) are representative of networks for which recording of signals is likely, while recording signals is unlikely in Fig. 2(a), (b), (e).

IV. MANIPULATING FIRING RATES WITH ASTROGLIA AND OPTIMIZED MEDIA

Although our cultures are virtually free of astroglia when initially plated [54], astroglia grow rapidly thereafter [55]. By our counting at one week the ratio of glia to neural cells is one after one week, increasing to a small multiple after several weeks, mimicking natural conditions [56]. If the neurons follow patterns, then the astroglia follow the neurons [Fig. 3(a)] [57], even though astroglia are much less likely to follow surface chemical cues than are neurons. We find that the emergence of widespread spontaneous electrical activity (one to two weeks) is correlated with the emergence of astroglia. We note that glia are known to be beneficial to this activity in vivo [58], [59].

Fig. 3.

Designing with astroglia and defined media. (a) Astroglia develop after neurons in culture, but much prefer neurons to cytophobic substrates, adding stability to a complex network (26 days in culture. Red astroglia, green neurons; reproduced with permission of Koninklijke Brill NV [57]). (b) Deliberately adding astroglia increases firing rates (reproduced with permission of Cambridge University Press [60].) (c) Deliberately fine tuning the media also increases neural activity. (reproduced with permission of Elsevier [61].)

From a signaling perspective, neural cultures have limited action potential firing rates, making it difficult to detect down-regulation of firing as part of a coding strategy. Adding extra glia to the cultures increases firing rates [Fig. 3(b)], prevents desensitization due to glutamate, causes less release from inhibition due to bicuculline, and likely increases the size of the inhibitory population [60]. Activity can also be increased by manipulation of the composition of the cell culture medium [61]. These cultures also exhibit considerable bursting that is suggestive of seizure activity. Stimulating the culture at moderate rates (e.g., 20 Hz) disrupts bursting phenomena and permits use of a wider range of stimuli for probing the network’s properties [62]. Hence the neuroengineer has several tools—glial addition, changes in media composition, or stimulation—with which to manipulate baseline activity level.

V. MICROTUNNELS—MEMS CAMPENOT CHAMBERS

Campenot chambers, reported in 1977 [63], consist of two neural growth compartments separated by a septum under which needle scratches defined paths for axonal extension from one compartment to the other. The modern MEMS equivalent was reported by Jeon’s group in 2005 [46] consisting of microfabricated tunnels in PDMS which is bonded to a substrate. In time the neurons in one compartment extend their axons through the tunnels, providing pure axonal material on which to do biomolecular analyses [64]–[66].

When combined with electrode arrays, the approach offers many substantial advantages. As shown in our work [47], the microtunnel PDMS structure is compatible with electrode arrays. Fig. 4 illustrates the concept. As shown in Fig. 4(b), the signals recordable from the axons in the tunnels are much larger (~100 μV) than in the open chambers where they are often so small as to be in the noise. This is due to the large series resistance of the narrow tunnel (~ 16 MΩ).

Fig. 4.

Microtunnel electrode array. (a) Schematic diagram showing four culture wells, each connected by a series of microtunnels to a central well. There are stimulating/recording electrodes in the wells and crossing the tunnels. The tunnels are 3 μm high by 10 μm wide by 750 μm long. (b) The amplitudes of the axonal potentials are much higher than normal due to the restricted and hence high resistance space in which the axons grow. Direction of propagation can be determined by the temporal order in which spikes occur on the electrodes place along the tunnels. (Reproduced by permission of The Royal Society of Chemistry (RSC) [47].)

The construct offers tremendous advantages for brain on a chip design. Each chamber can be filled with a different population of neurons (e.g., granule or pyramidal cells) or of muscle fibers. Unidirectional transmission can be created through timed growth of the neurons. Various logical geometries of network can be constructed. Separate fluidic treatment of the different populations should be possible. The signals from within the tunnels provide a unique opportunity to sample action potential signals from which one can separately monitor the activity of the source, the target, and the communication channel between them.

VI. BRAIN IN A CHIP

The future will include three-dimensional constructs. There has been progress in developing culture techniques to support neural growth in three-dimensional hydrogels [67]–[69], with experiments showing the critical importance of fluid flow within the hydrogels to enhance cell survival [70]. We and others [71], [72] have printed or photolinked neural guidance molecules onto deformable substrates. These suggest that it will be possible to create three-dimensional substrates in which neurons grow in controlled geometric patterns.

Recording signals from three-dimensional constructs will be challenging. Already two designs have appeared in the literature (Fig. 5; [73], [74]). Both offer multiple electrode contacts and perfusion ports for either maintenance of tissue or the application of drugs for testing. The layered structure proved especially compatible with existing commercial hardware as electrode contacts could easily be made to match the footprint of commercial amplifiers. Demonstration experiments showed that neural cultures could be kept alive for weeks, that they developed correlated spontaneous activity and could be stimulated. Further fluidic perfusion and drug application were successfully performed. Hence these offer models for further development of “braininachip” technologies.

Fig. 5.

3-D microelectrode arrays for neural culture. (a) Picket fence style 3-D electrode array with integrated electrodes and fluidics [74]). (b) Layered array used successfully to culture neurons, record correlated activity on different layers, and deliver drug to alter activity [73]). (c), (d) views of neurons growing on the top and middle locations in the chamber. [All figures reproduced by permission of The Royal Society of Chemistry (RSC).]

VII. CODING AND PLASTICITY

There is a substantial literature investigating how information may be represented in patterns of action potential activity recorded from MEA, and only a suggestion as to the breadth of the studies is possible here. Strong model formation has come from physicists [75], including modeling bursts as avalanches [76], the use of information theory [75] and affinity and field theory measures [77], [78], and engineers with state-space [79], clustering and multidimensional scaling approaches. Characteristic bursting events [80]–[82]} are often the focus of the analysis, including analyses of burst propagation [83]. However, much is to be done because of the high dimensionality of the data (often 60 or more channels of activity) and wide spread in the time scales of importance (submillisecond for synaptic phenomena, weeks for developmental phenomena).

Progress is being made in developing learning and memory models. Plasticity in burst response has been quantified [84]. The work of Jimbo et al. [85], recently repeated with Grainger causality analysis [86], showed widespread but highly variable potentiation and depression of the pathways connecting stimulating electrode to recorded neuron. Ruaro et al. showed that elementary pattern recognition and signal processing functions could be impressed on a cultured network [87]. Marom’s group showed differential learning of rare and frequent stimuli [88], which formed the basis for DeMarse’s demonstration of control of a simulated device [89]. Feedback control of network behavior has been demonstrated [90].

VIII. FUNCTION FOLLOWS FORM

We are seeing the beginning of the exploration of how functional properties are affected by network geometry. We see increases in neural activity with chronic stimulation of cultured networks [91] and increases in activity with synapse density [92]. These predict enhanced development over weeks of networks geometrically patterned to have greater connectivity, for which we have preliminary evidence [93]. Different sized networks show different statistical properties [94]. A more dramatic example of how geometrically designed networks can have designed function is the narrow line (150 μm) network of Jacobi and Moses [95] wherein propagation of electrical activity is a function of cell density. More recently Feinerman et al. showed how to design the functional equivalents of delay lines, diodes, and and gates with patterned neurons [96].

IX. DISCUSSION AND SUMMARY

The goal of creating a brain on a chip has certainly not been met. However, the progress toward controlling and understanding neural activity in a culture dish has come a long way in the thirty years since first started, and the less than twenty years since the first serious attempts to create patterning technologies. As highlighted above, there are many encouraging reports of progress toward not only understanding the behaviors of these in vitro networks, but of being able to customized their functional properties so that they may be more valuable in the pursuit of basic and applied science applications.

The report above does not highlight the difficulties in performing the work. While culturing, recording and stimulating neurons without patterns is routine in many laboratories around the world, the addition of patterning greatly increases the complexity and risk and reduces the likelihood of any one culture being successful. Still, in our hands networks similar to those in Fig. 2(c), (d) maintain reasonable fidelity to their patterns for a month often enough for experimentation; some have survived for several months with relatively good fidelity [51].

The retreat from a focus on networks of single neurons (from, e.g., [97], [98]) appears largely due to these difficulties, especially for mammalian preparations, leading to “spoke-and-cluster” or “street-plan” constructs for which activity is more robust. Migration to Campenot chamber type recordings with distinct populations of cells in the different wells will be a natural response to searching for a more robust experimental platform. The addition of glia and change of media represent attempts to address the issue of robust electrical activity.

A fortunate aspect of this work has been the tremendous advance in instrumentation and signal processing available from commercial vendors and from software freely available on the web. The front end processing—detecting signals, averaging, performing pairwise correlations, and extracting basic signal properties—is easily accomplished. (We note, however, that there is growing development of integrated circuit microelectrodes with thousands of channels for stimulation and recording, promising a substantial shift in how researchers approach neural culture studies. [99]–[101]) The current technology has enabled investigators to pursue sophisticated statistical approaches to understanding the underlying activity. Much remains to be done, however, as the dimensions of the data set and the inherent nonstationarity of the preparation provide great challenges.

X. EXPERIMENTAL OVERVIEW

Our work has involved mostly primary rat hippocampal or cortical cells taken from embryonic day 18 rat pups. They are cultured according to standard protocols that may be found in the referenced papers. Briefly, cells are taken from the embryonic brains, dissociated, and allowed to settle onto the electrode array or culture dish surfaces, attaching as shown in multiple figures in this paper. They can be maintained for several months if not longer. All procedures were approved by animal use protocols at the University of Illinois and at the Southern Illinois University School of Medicine.

Biography

Bruce C. Wheeler (Fellow, IEEE) received the B.S. degree from the Massachusetts Institute of Technology, Cambridge, and the M.S. and Ph.D. in electrical engineering from Cornell University, Ithaca, NY.

He was at the University of Illinois for 28 years, including service as Founding and Interim Head of the Bioengineering Department, Associate Head of Electrical and Computer Engineering, and Chair of the Neuroscience Program. He was Professor of BioE, ECE, and the Beckman Institute. He recently moved to the University of Florida and was appointed Interim Chair of the J. Crayton Pruitt Family Department of Biomedical Engineering. His research interests lie in the application of electrical engineering methodologies to neuroscience, including neural spike sorting technologies, microelectrode arrays for use with brain slices and cell culture, lithography to the geometric patterns of neural growth in culture.

Prof. Wheeler is a Fellow of the AIMBE. He is the Editor in Chief of the IEEE Transactions on Biomedical Engineering.

Gregory J. Brewer received the B.S. degree from California Institute of Technology, Pasadena, and the Ph.D. degree in biology from the University of California, San Diego.

He holds the Kenneth Stark Endowed Chair in Alzheimer’s Research and is Professor of Neurology and Medical Microbiology, Immunology and Cell Biology at Southern Illinois University School of Medicine, where he has been for 29 years. His research interests lie in the application of engineering and biochemistry principles to neuron cell culture, development and live neural networks. He also has projects on the neuroscience of aging and is founder and President of BrainBits LLC, which supplies live brain tissue to neuroscientists.