Session A11: Physics of Neural Systems I

In vitro living neurons in a reservoir computing framework

The Reservoir Computing (RC) paradigm has been devised for the processing of dynamic, real-time, and multimodal input signals. The core mechanism of RC relies on decoding output signals from the reservoir in a supervised manner, while the reservoir itself serves the purpose of  “digesting” inputs recurrently. This leads to computationally efficient machinery that does not require expensive backpropagation weight updates, as in classical recurrent neural networks (e.g. RNN/LSTM). Nonetheless, the choice of the reservoir is critical, and a good reservoir requires the ability to express complex dynamics with sufficient non-linearity and separation properties. In this work, we target live in vitro neuron cultures, known for their scalable hypernetworks and complex spiking dynamics, making them potentially suitable for constructing quality reservoirs. Additionally, the self-organizing and malleable nature of neuronal synapses may allow the biological neural network to adapt to problems continuously and more quickly.

To build a living reservoir from neuron cultures, we integrate optogenetic stem cell-derived neurons with micro-electrode arrays (MEA). Electrical and optical stimulation are employed to feed encoded information into the biological neural networks (BNN), and the reservoir state is represented by spiking activities. We exploit the reservoir capability of BNNs and demonstrate them through benchmark problems.   

Collective Dynamics of Astrocytes

Astrocytes are the most abundant cell type in the brain. While they are not capable of rapid depolarization, they exhibit dynamics in their calcium levels and regulate the information flow in the brain by coupling to each other via calcium oscillations and to neurons via the Tri-Partite synapse. Here we analyze the collective calcium dynamics of astrocytes in three distinct states: (1) stellate, (2) polygonal, and (3) reactive that correspond to immature, mature, and injured astrocytes, respectively. We analyze various metrics of collective network activity of astrocytes, and assess how collective behavior adapts to differing morphologies and changing chemical environments.   

A versatile, accurate, low-cost and open-source electrophysiology solution for in-vitro neuronal cultures

We present a versatile, accurate, and low-cost in-vitro electrophysiology solution based on open-source electronics, fully customizable recording platforms, and in-house micro-electrode arrays (MEAs) substrates. The recording platform is made of standard inexpensive materials and can accommodated a variety of MEAs configurations entailing up to 512 recording channels. Built-in modularity allows for the seamless integration of add-on modules such as electrical/optical stimulation devices, as well as on-board sensors and fluidic interfaces. MEAs fabrication relies on maskless photolithography, greatly favoring rapid prototyping, here demonstrated in a variety of patterns involving different electrode densities, diameters, and spatial topologies. Through a native open-source data acquisition terminal and analysis software, we demonstrate the functionalities of our system across a range of neuronal cultures, from embryonic stem cell-derived neurons and primary brain cells to organotypic brain slices. This approach significantly lowers barrier entry to electrophysiology studies via ten-fold cost reductions, while providing a flexible hardware solution that can be widely deployed.

    

A functional connectivity atlas of C. elegans measured by neural activation

Neural processing and dynamics are governed by the details of how neural signals propagate from one neuron to the next through the brain. We systematically measured functional properties of neural connections in the head of the nematode Caenorhabditis elegans by direct optogenetic activation and simultaneous calcium imaging of 10,438 neuron pairs. By measuring responses to neural activation, we extracted the strength, sign, temporal properties, and causal direction of the connections and created an atlas of causal functional connectivity. We find that functional connectivity differs from predictions based on anatomy, in part, because of extrasynaptic signaling. The measured properties of the connections are encoded in kernels which describe signal propagation in the network and which we fit from the data. Using such kernels, we can run numerical simulations of neural activity in the worm's brain using exclusively information that comes from the data, as opposed to simulations based on the anatomical connectome which require assumptions on many parameters. We show that functional connectivity better predicts spontaneous activity than anatomy, suggesting that functional connectivity captures properties of the network that are critical for interpreting neural function.   

Olfactory learning and navigation behavior of C. elegans in a controlled odor environment

Animals flexibly adjust behavior in response to sensory environments after learning experiences. In the nematode worm C. elegans, associative learning with an olfactory cue modulates chemotactic behavior towards the cue if it was paired with food or starvation. However, it is unknown how olfactory learning alters sensory-motor processing for any specific navigation strategy. For example, whether it adjusts the bias in a biased random walk, or alternatively modulates other navigational strategies adaptively during the task. The biophysics with which worms sense airborne olfactory cues is also not well understood. Here we investigate butanone-odor associative learning by constructing a flow chamber to precisely measure the odor concentration experienced by worms during odor-guided navigation. We control airborne cues to form a stationary chemical landscape, monitor concentration with an array of digital odor sensors, and track the trajectories and posture of worms in the environment. We developed a statistical model with a mixture of navigation strategies to characterize chemotaxis behavior. This model captures different strategies previously reported in worms, including biased random walk (klinokinesis) and gradual change in the head angle towards high concentration (klinotaxis). With our apparatus and proposed model, we will discuss progress towards quantitatively characterizing the bi-directionally modulated navigation strategies in learned odor-guided navigation.   

Touch inhibits feeding via a neural bottleneck in C. elegans

A neural bottleneck is characterized by the projection of multiple neurons onto a smaller

number of neurons. This network architecture suggests that the information encoded in the incoming signals is compressed at the bottleneck. The nervous system of the nematode C. elegans features such a bottleneck motif. The 302 neurons that make the brain of this animal are separated into two sub-networks - the somatic and the pharyngeal networks - that are only connected by a pair of gap junctions. Sensory information, for instance mechanical stimulation, recorded at the somatic network travels through the bottleneck and suppresses the pumping motion of the pharynx. To investigate the role of this bottleneck in information coding, we supply controlled mechanical stimuli to C. elegans while observing changes in feeding behavior.

Estimating information compression requires a large amount of measurements of both the input

and output signals. Therefore, we implemented a high-throughput assay to supply substrate vibrations with tunable amplitude and frequency as mechanical stimulus while observing pumping in C. elegans populations. Using our custom image analysis pipeline ‘PharaGlow’, we can detect automatically pumping events in multiple animals moving on standard cultivation plates (Bonnard and Liu, eLife 2022). Consistent with previous studies, we find that vibrations inhibited pumping and that touch information flows through the bottleneck. We will discuss the utility of this model system to study sensory information encoding and compression.   

Turning-associated neurons gates a sensorimotor response in C. elegans upstream of neuron AVA

A fundamental task of the brain is to integrate sensory cues with an animal’s current behavior to respond appropriately. But precisely how neural circuits integrate sensory and behavior signals remains unknown. To address this, we investigated C. elegans response to touch. C. elegans often respond to a mechanosensory stimulus by moving backward, called a reversal. Previously we found that C. elegans are less likely to reverse in response to mechanosensory stimuli delivered when it is in the midst of turn compared to stimuli delivered during forward locomotion. To explore the underlying circuit mechanisms, we optogenetically activated downstream interneurons in the mechanosensory pathway, either while the worm was moving forward or while it was turning. Activation of interneurons AIZ, RIM, AIB, or AVE,  during forward locomotion was more likely to evoke reversals than during turning. Activation of neuron AVA, however, evoked reversals with similar likelihood regardless of whether the animal was moving forward or turning. Further, inhibiting turning-related neurons RIV, SMB, and SAAD, restored the animal’s response to mechanosensory stimuli, even during turns. Taken together, our measurements suggest a circuit mechanism in which signals from turning associated neurons act as a gate to disrupt or inhibit mechanosensory-related signals, and that these turning related signals act upstream of neuron AVA.   

Statistical Significance Analysis of Functional Connectivity Measurements of the Brain

New optical methods for measuring and manipulating neural activity in the brain promise unprecedented insights into how signals propagate through a biological neural network at brain scale and single neuron resolution. However, the scale of these large-scale measurements poses challenges for interpretation. We recently measured how neural signals propagate in the nematode Caenorhabditis elegans by systematically activating individual neurons in the head and measuring the neural network's response at cellular resolution. The resulting dataset contains 10,438 measurements of neural responses to stimulation measured across 43 animals. The scale of the measurements requires a rigorous statistical framework in order to exclude apparent neural responses that could likely be attributed to random chance. Here we discuss a statistical framework for accurately conveying our confidence in attributing one neuron's response to the activation of another.   

Adaptive robustness through incoherent signaling mechanisms in a regenerative brain

Animal behavior emerges from collective dynamics of interconnected neural populations, making it vulnerable to damages to the connectome architecture. However, many organisms can maintain significant behavioral output in the face of large-scale neural damages, though molecular underpinnings of this extreme robustness remain mostly unknown. Here, we develop a high-content imaging platform and quantitative behavioral analysis framework that enable us to measure a previously uncharacterized long-lasting latent memory state in planarian flatworms during their whole-brain regeneration with unprecedented precision. By combining over 10,000 animal trials with computational modeling of neural population dynamics, we show that long-range volumetrically transmitted peptidergic signals allow the nervous system to maintain robust control over behavior when large portions of neurons are ablated. The different time and length scales of neuropeptide and small molecule transmission lead to incoherent signals competitively regulating the latent memory. During regeneration, long-range peptide signals dominate, allowing the planarian to rapidly reestablish the latent behavioral state and restore coarse behavioral output while gradually refining it to a precise response. Controlling behavior through two overlapping but opposing communication mechanisms creates a more robust system than either alone and may serve as a generic approach to construct robust neural networks.   

Network Properties the whole-brain Drosophila connectome

Characterizing the network properties of animal brains may lead to a better understanding of computation and information flow in these complex organs. However, to date very few whole-brain neuron-level reconstructions have been completed across organisms. The FlyWire project has completed the proofreading of a connectome for a Drosophila female brain (FAFB) which contains both complete hemispheres of the central brain and includes neurons that receive inputs in the optic lobes. Here, we dissect the network properties of the complete central brain of the fruit fly. We characterized the distributions of synaptic connection weights and network motifs in 75 anatomically defined brain regions, or neuropils, and found that different neuropils have different network statistics. We constructed a projectome describing how connected these neuropils are to each other, and identified the dominant neurotransmitters exchanged between each neuropil pair. To group neurons by connectivity rather than by anatomy, we employed spectral clustering to sort neurons into modules. We found that many of these modules correspond to neuron classes with known biological functions. Finally, we identified groups of neurons which project across the midline, a population which likely plays a critical role in sending signals from one hemisphere of the brain to the other. Together, these results highlight how the topology of the anatomical neuronal network may direct and constrain the flow of information across the brain of the fly.   

Coordinated neural activity in freely behaving Drosophila larvae

To understand how brains encode and coordinate behaviors, we seek to follow neural activity in organisms behaving in naturalistic manners. Using volumetric, multi-neuronal imaging in freely behaving Drosophila larvae, we measure neural activity from an extended volume within the larval central nervous system. Recordings from motor and premotor neurons show waves of activity propagating in sync with a larva’s peristaltic crawling, thus demonstrating the power of our microscope to aid in answering this long-standing question in systems neuroscience.   

Neural Navigation: rapid learning in complex partially observed environments

Brains arguably evolved in order to coordinate the movement of organisms in space. Much is known about how sensory stimuli are transduced into neural signals and how brains generate motor actions. However, it remains unclear how brains convert sensory and internal signals into complex sequences of actions. How do neural tissues switch between different modes of action, such as exploration of the environment vs exploitation of known direct routes to rewarding locations? How are navigation-related error signals computed? What neural constraints govern effective navigation through a complex environment? Traditionally, these questions have been difficult to address in animal subjects, due to the long training times required for even simple tasks. Here, I leverage the natural propensity of rodents to explore and memorize complex environments. Experiments in complex mazes allow multi-action sequences along direct routes between locations of interest to be learned in a single experimental session. Complementary modeling approaches, drawing inspiration from physics and machine learning, help to characterize the range of possible biological solutions to these problems and to generate predictions for the neural activities recorded from navigating rodents.   

A family of neurodevelopmental biomarkers extracted from a statistical analysis of kinematic data measured with high-definition sensors

There is significant interest in finding quantitative ways of diagnosing psychiatric disorders. Most clinical studies use qualitative assessments, based on clinical interview and behavioral observations. In our research program, we have concentrated on studying subjects with neurodevelopmental disorders under the hypothesis that significant diagnostic information is contained within how people move. We used as an experimental paradigm, targeted reaching movements and examining kinematic variables at millisecond time scales. In our previous studies [1,2] of subjects with Autism Spectrum Disorder (ASD) , we analyzed the randomness of maximal velocity. We reported that the probability distribution for its maximum can be mapped directly to ASD diagnosis. We now extend these studies of velocity randomness to participants with Attention Deficit Hyperactive Disorder (ADHD), and comorbid ASD+ADHD [3]. We identified a new family of biomarkers related to the ratio of an overlap of distribution areas under their curves.