Neuroscience 2021

Technological advances such as single-cell RNA sequencing accelerated our understanding of cellular diversity in tissues. However, the ability to elucidate this cellular heterogeneity while retaining spatial information is crucial for understanding the complex biological networks in health and disease. At Resolve Biosciences we developed Molecular CartographyTM, a highly multiplexed, single-molecule detection technology to measure transcriptomic activity with full spatial context at subcellular resolution. The unparalleled sensitivity and specificity of Molecular CartographyTM helps scientists detect individual transcripts and rare signals to interpret fundamental biology and rapidly advance the understanding of complex biological questions in neuroscience. In this presentation, we demonstrate applications in mouse and human brain tissue in which we detect millions of individual transcripts originating from a hundred genes in each sample. This allows us to unravel the complex spatial organization of different cell types and their changes in gene expression in response to pathology, for example, in the context of Alzheimer's disease.

Learning Objectives:

1. Method for highly multiplexed and sensitive spatial transcriptomics

2. Spatial analysis of cell phenotypes in response to pathology in the context of Alzheimer’s disease

In 1910, Harrison published the first report of frog embryonic sympathetic ganglia grown in hanging drops of lymph for a few days, where single neurons extended nerve fibers with complex growth cones extending and contracting over time towards target cells. The revolutionary ability to culture embryonic and post-natal neurons for extended time periods demonstrated that developing neurons could form plexi of interconnected nerve fibers. However, over the years it has proven to be exceedingly difficult to consistently culture adult central nervous system (CNS) neurons; while several instances of success in culturing fewer than 10 neurons have been reported, a major unresolved methodological challenge in neuroscience has been the ability to culture adult CNS neurons from any region consistently and in tractable numbers. It has been possible to culture adult dorsal root ganglia neurons for reasons that remain unclear; these neurons also retain an ability to regenerate throughout life. The value of CNS neurons in advancing understanding of intact adult neuronal function, response to injury, and as a screening system for candidate therapeutics is clear. Here we report for the first time the ability to consistently culture adult neurons of specified phenotype in large numbers over extended time periods. We further provide evidence for the utility of these cultures in advancing studies of CNS injury.

Learning Objectives:

1. Successful culture of adult primary neurons from the central nervous system

2. Electrical recordings and stimulation of neural networks in vitro

3. Practical applications for primary adult CNS cultures

Multiscale modeling has arisen as a focus of computational systems biology, with the realization that genome, proteome, connectome, etceteromes, will only become comprehensible once placed in the context of explicit computer simulations. Measurements and activity patterns at one scale must be understood dynamically in the context of patterns at higher and lower scales. Nowhere is this more apparent than in the domain of brain disease, where manipulations at molecular levels (drugs) are used to change, licitly or illicitly, behavior and  thought.  We have begun to study schizophrenia and other brain dysfunctions with an eye towards connecting these levels in order to reveal functional and dysfunctional dependencies. Rather than only thinking hierarchically up the great chain of embeddings (e.g. from molecule to spine to dendrite to cell to circuit to area to behavior), it will be valuable to transform to other representational frames to connect across multiple scales.

Learning Objectives:

1. -Distinguish approaches, objectives, synergies of 1. mechanistic multiscale modeling; 2. phenomenological modeling; 3. AI through deep learning.

2. Appreciate how genomic, molecular, transcriptomic thinking and results can be integrated into functional contexts at higher scales.

3. Assess how datamining and simulation techniques interface with experiment and help integrate experimental measurements across scales.

One role of theory is in guiding future experiments: What should we aim to measure? Which experimental results should we be surprised about? I will argue here that simple random networks models of neural dynamics explain, to the experimental accuracy, many features of large population recordings in a variety of nervous systems. Observation of such features is thus not surprising in the context of random null models. I will discuss how this affects, or can affect, our interpretation of experiments, our choices of which experiments to do next, and how it raises hopes for building next-level theories of nervous systems, finally leaving the neuron behind.

Learning Objectives:

1. Explain how large simple neural networks can result in complex experimental recordings.

2. Introduce the concept of latent variables dynamics

3. Define coarse-graining modeling approaches

Recent advances in machine learning have shown that deep neural networks (DNNs) can provide powerful and flexible models of neural sensory processing. In the auditory system, standard linear-nonlinear (LN) models are unable to account for high-order cortical representations, but thus far it is unclear what additional insights can be provided by DNNs, particularly in the case of single-neuron activity. DNNs can be difficult to fit with relatively small datasets, such as those available from single neurons. In the current study, we developed a population encoding model for a large number of neurons recorded during presentation of a large, fixed set of natural sounds. Leveraging signals from the population substantially improved performance over models fit to individual neurons. We tested a range of DNN architectures on data from primary and non-primary auditory cortex, varying number and size of convolutional layers at the input and dense layers at the output. DNNs performed consistently better than LN models. Moreover, the DNNs were highly generalizable. The output layer of a model pre-fit using one population of neurons could be fit to different single units and/or different stimuli, with performance close to that of neurons in the original fit data. These results indicate that population encoding models capture a general set of computations performed by auditory cortex and can be analyzed to derive a characterization of auditory cortical function. Tools developed for this project as part of the BRAIN Initiative can be applied to a wide range of problems in the auditory and other neural systems.

Learning Objectives:

1. Applying deep learning to neural populations overcomes data limitations that hinder analysis of single-neuron data.

2. Deep learning models consistently outperform traditional linear-nonlinear models of single neurons in auditory cortex.

3. Population-based models generalize well to neurons not included in the original model fit.

Cortical neurons process information on multiple timescales, and neurons that remain active during delay periods lasting seconds have been recorded in the parietal and prefrontal cortex.  However, the underlying mechanisms for the emergence of neuronal timescales stable enough to support working memory are unclear. A spiking recurrent neural network (RNN) model was trained on a working memory task and unit activity was compared with single neurons in the primate prefrontal cortex.  The temporal properties of our model and the neural data are remarkably similar. Analysis of the RNN model revealed strong inhibitory-to-inhibitory connections underlying a disinhibitory microcircuit as a critical component for long neuronal timescales and working memory maintenance. Enhancing inhibitory-to-inhibitory connections led to more stable temporal dynamics and improved task performance. The same network can perform other tasks without disrupting its pre-existing timescale architecture, suggesting that strong inhibitory signaling underlies a flexible working memory network.

Learning Objections:

1.  To learn a new approach to understanding brain function by creating networks to perform tasks using machine learning

2.  To learn how how to probe these networks to discover how they solve a task

3.  To learn how these models can be used to formulate experimentally testable hypotheses

Cortical circuits often receive multiple inputs from upstream populations with non-overlapping stimulus tuning. Both the feedforward and recurrent architectures of the receiving cortical layer reflect input tuning diversity. We study how population-wide neuronal variability propagates through a hierarchical cortical network receiving multiple independent tuned inputs. Our new analysis of in vivo neural data from the primate visual system shows that the number of latent variables (dimensions) needed to describe population shared variability is smaller in V4 than in downstream visual area PFC.  We successfully reproduce the dimensionality expansion observed across brain areas using a multi-layer spiking network with structured feedforward projections and recurrent assemblies of multiple, tuned neuron populations. We show that tuning-structured connectivity generates attractor dynamics within the model PFC activity, where attractor competition expands the dimensions of shared variance. Restricting dimensionality analysis to activity from one attractor state recovers the low-dimensional structure inherited from each of our tuned inputs. Our model thus introduces a framework where high-dimensional cortical variability is understood as alternating states of low-dimensional, tuning-specific circuit dynamics.

Learning Objectives:

1. Measure the space in which population-wide neuronal fluctuations inhabit in both visual area V4 and the prefrontal cortex (PFC).

2. Build a circuit based model that captures how neuronal variability is transferred from V4 to PFC.

3. Using our model, build a theoretical framework that predicts how the recurrent circuitry within PFC supports an expansion of the dimensions of variability compared to V4.

Targeted stimulation of the brain has the potential to treat mental illnesses but designing an appropriate protocol requires a multitude of choices.  I will describe an approach to help design the stimulation protocol by identifying electrical dynamics across many brain regions that relate to illness states or behaviors, defined as electrical connectome, or electome, networks.  The observed electrical activity of the brain is modeled as a superposition of activity from the latent electome networks, where the weights on the latent networks predict disease state, behavior, or outcomes.  These electome networks are interpretable in their power and coherence relationships between brain regions, forming an explainable artificial intelligence approach for neural data that facilitates future experimental design.  I will highlight how this approach has been used to successfully design stimulation protocols in the context of an animal model of major depressive disorder and appetitive social behavior.

Learning Objectives:

1. Define an additive model

2. Explain how decompositions of neural activity can be linked with disease states

The intrinsic activity of the brain is organized into networks and motifs that vary over time. To understand how coordinated macroscale patterns of intrinsic activity flow across the brain’s structural network, new analysis approaches that can extract interpretable spatial and temporal features are required.  We demonstrate the use of machine learning to extract prototypical spatiotemporal trajectories of brain activity that provide a sparse and elegant basis of representation. To better interpret the spatiotemporal patterns of brain activity, we incorporate brain network models as constraints, forcing the algorithm to learn parameters that can be measured empirically. This interpretable machine learning-based approach paves the way for multimodal experiments that can validate or disprove the accuracy of different brain network models.

Learning Objectives:

1. Define brain network models and explain their application in the study of macroscale activity in the brain.

2. Identify machine learning approaches to characterizing brain dynamics.

3. Describe the use of brain network models to constrain machine learning techniques and improve interpretation.

We will show how to combine large scale neural recordings and mechanistic neural network models to advance our conceptual understanding of how neural circuits mediate cognitive functions like perception, navigation and semantic cognition.  With additional mathematical analysis, we can also even explain why some neural circuits might be organized the way they are.  First, we will describe state of the art models of the retinal response to natural movies, and use explainable artificial intelligence to understand how they recapitulate over 20 years of retinal physiology experiments.  Second we will describe how the 4 most dominant cell-types in our retina emerge naturally as a consequence of optimal spatiotemporal processing of natural movies.  Third, we will demonstrate how hexagonal grid cell firing fields must obligatorily arise in any biologically plausible neural network that is capable of performing path-integration.   And fourth, time permitting, we will describe how the hierarchical differentiation of concepts that unfolds over time in infant semantic development can be accounted for by simple mechanistic neural models. 

Learning objectives:

1. Describe how to create mechanistic neural models of cognitive phenomena

2. Understand explainable artificial intelligence and how to use it to extract conceptual insights from mechanistic models

3. Learn about optimality principles for explaining why some neural circuits are organized the way they are

4. Learn about specific phenomena in perception, navigation, and semantic cognition
 

Understanding how populations of neurons work together to represent stimuli, build percepts, and generate complex behaviors, is a fundamental challenge in neuroscience. To establish a link between neural activity and behavior, especially in complex behaviors and tasks, we need approaches that allow for an unbiased look into the latent factors that underlie neural activity. In this talk, I will describe ways in which my lab is tackling this objective through the development of novel unsupervised representation learning methods. With new tools to read out information from the brain in a stable and generalizable manner, we can enable rich comparisons across neural recordings spanning different time points in learning, or across aging and development.

Learning Objectives:

1. Audience members will learn about new tools for visualizing and interpreting neural activity

2. Learn about the application of representation learning methods on various neural recordings from macaque, rat, and mouse

A central challenge in neuroscience is to decipher how cellular mechanisms are coordinated across multiple aspects of the circuits that control whole-animal behavior. We will briefly discuss how modern neuroscience can strive to meet these goals. A key challenge is to unify a top-down approach to brain, computation and behavior, as defined by David Marr and Horace Barlow, with a bottom-up approach that takes inspiration from genes and genetic models, molecules and cell biology. Both of these viewpoints have merit of course, and both find common ground and opportunity for unification at the level of brain circuits. The story behind this talk –too young to be declared a success – concerns a multi-PI NIH U-19 project titled Oxytocin Modulation of Neural Circuit Function and Behavior. My colleagues Robert Froemke (co-PI), Dayu Lin and Gyorgyi Buzsaki and I have focused our labs’ attention on the peptide modulator oxytocin, the neural circuits it affects, and its role in orchestrating mouse social behavior, critical for survival yet still incompletely understood. Oxytocin signaling proves essential for the social transmission of maternal behavior (Carcea, Froemke et al. Nature 2021) and vital for acquisition and storage of information about social rank (Osakada, Lin et al. in preparation). With some striking social behaviors as a starting point we seek clarification of “socio-spatial memory neural circuits” and their operation in disparate brain regions such as auditory cortex, hippocampal area CA2, lateral septum and VMHvl of the hypothalamus. From an evolutionary perspective, work on oxytocin addresses the general question of how neuromodulators can transform circuits by operating at a different level than fast neurotransmitters. It also contributes to a wider effort to comprehend “the molecules, cells, and circuits linking sensory information to the motor outputs of social behavior”, which Park and Insel called “the dark matter” of social neuroscience. In so doing, we cannot help but evaluate the challenges of multilevel, multi-investigator projects, how synergies can be sought and fostered, and indeed how success itself might be defined.

Learning Objectives:

1. Delineate the goals of studies of brain, computation and behavior and how they can be integrated

2. Describe how neuromodulators such as oxytocin transform circuits and operate at a different level than fast neurotransmitters

3. Evaluate the synergies and challenges of multilevel, multi-investigator projects and how success could be defined

Complex behaviors are often driven by an internal model, which integrates sensory information over time and facilitates long-term planning to reach subjective goals. A fundamental challenge in neuroscience is: how can we use behavior and neural activity to understand this internal model and its dynamic latent variables? Here we interpret behavioral data by assuming an agent behaves rationally — that is, it takes actions that optimize its subjective reward according to its understanding of the task and its relevant causal variables. We apply a new method, Inverse Rational Control (IRC), to learn an agent's internal model and reward function by maximizing the likelihood of its measured sensory observations and actions. This thereby extracts rational and interpretable thoughts of the agent from its behavior. We also provide a framework for interpreting encoding, recoding and decoding of neural data in light of this rational model for behavior. When applied to behavioral and neural data from simulated agents performing suboptimally on a naturalistic foraging task, this method successfully recovers their internal model and reward function, as well as the Markovian computational dynamics within the neural manifold that represents the task. This work lays a foundation for discovering how the brain represents and computes with dynamic latent variables.

Learning Objectives:

1. Distinguish rational behavior from optimal behavior

2. Identify how to use estimates of latent mental variables to predict brain dynamics

Threshold-linear networks (TLNs) display a wide variety of nonlinear dynamics including multistability, limit cycles, quasiperiodic attractors, and chaos. Over the past few years, we have developed a detailed mathematical theory relating stable and unstable fixed points of TLNs to graph-theoretic properties of the underlying network. These results enable us to design networks that count stimulus pulses, track position, and encode multiple locomotive gaits in a single central pattern generator circuit.

Learning Objectives: 

1. What types of attractors that can be encoded in inhibition-dominated neural networks? Can multiple dynamic attractors coexist in the same network?

2. What neural network functions or computations can be performed by threshold-linear networks? Give two examples.