Using artificial-intelligence-driven deep neural networks to uncover principles of brain representation and organization
Host: C. Kirst
Human behavior is founded on the ability to identify meaningful entities in complex noisy data streams that constantly bombard the senses. For example, in vision, retinal input is transformed into rich object-based scenes; in audition, sound waves are transformed into words and sentences. In this talk, I will describe my work using computational models to help uncover how sensory cortex accomplishes these enormous computational feats.
The core observation underlying my work is that optimizing neural networks to solve challenging real-world artificial intelligence (AI) tasks can yield predictive models of the cortical neurons that support these tasks. I will first describe how we leveraged recent advances in AI to train a neural network that approaches human-level performance on a challenging visual object recognition task. Critically, even though this network was not explicitly fit to neural data, it is nonetheless predictive of neural response patterns of neurons in multiple areas of the visual pathway, including higher cortical areas that have long resisted modeling attempts. Intriguingly, an analogous approach turns out be helpful for studying audition, where we recently found that neural networks optimized for word recognition and speaker identification tasks naturally predict responses in human auditory cortex to a wide spectrum of natural sound stimuli, and help differentiate poorly understood non-primary auditory cortical regions. Together, these findings suggest the beginnings of a general approach to understanding sensory processing the brain.
I'll give an overview of these results, explain how they fit into the historical trajectory of AI and computational neuroscience, and discuss future questions of great interest that may benefit from a similar approach.
From active liquid crystals to de-wetting liquid droplets: Using mesoscopic models to understand the physical behaviors of cells and tissues
Host: T. Shendruk
Living cells generate and transmit mechanical forces over diverse time-scales and length-scales to determine the dynamics of cell and tissue shape during both homeostatic and pathological processes, from early embryonic development to cancer metastasis. These forces arise from the cell cytoskeleton, a scaffolding network of entangled protein polymers driven out-of-equilibrium by enzymes that convert chemical energy into mechanical work. However, how molecular interactions within the cytoskeleton lead to the accumulation of mechanical stresses that determine the dynamics of cell shape is unknown. Furthermore, how cellular interactions are subsequently modulated to determine the shape of the tissue is also unclear. To bridge these scales, our group in collaboration with others, uses a combination of experimental, computational and theoretical approaches. On the molecular scale, we use active nematic liquid crystals as a framework to understand how mechanical stresses are produced and transmitted within the cell cytoskeleton. On the scale of cells and tissues, we abstract these stresses to surface tension in a liquid droplet and draw analogies between the dynamics of droplet wetting (and dewetting) and the shape dynamics of cells and simple tissues. Together, we attempt to develop comprehensive description for how cytoskeletal stresses translate to the physical behaviors of cells and tissues with significant phenotypic outcomes such as cancer metastasis.
The brain builds its internal sensory representations based on the structure of neural activity. A number of modalities, such as the early visual system and the spatial map in hippocampus, are relatively well-characterized. However some sensory systems, such as olfaction, remain enigmatic. Perhaps the primary difficulty is that the underlying perceptual space is not well-understood. Can we "build" the sensory space from neural activity alone, without a prior understanding of how the stimuli are organized? I will describe a set of mathematical tools that allow to infer the dimension and the topology of stimulus space and illustrate their utility for two neural systems: hippocampus and early olfaction.
Theory of margination in blood and other multicomponent suspensions
Host: T. Shendruk
Blood is a suspension of objects of various shapes, sizes and mechanical properties, whose distribution during flow is important in many contexts. Red blood cells tend to migrate toward the center of a blood vessel, leaving a cell-free layer at the vessel wall, while white blood cells and platelets are preferentially found near the walls, a phenomenon called margination that is critical for the physiological responses of inflammation and hemostasis. Additionally, drug delivery particles in the bloodstream also undergo margination – the influence of these phenomena on the efficacy of such particles is unknown.
In this talk a mechanistic theory is developed to describe segregation in blood and other confined multicomponent suspensions. It incorporates the two key phenomena arising in these systems at low Reynolds number: hydrodynamic pair collisions and wall-induced migration. The theory predicts that the cell-free layer thickness follows a master curve relating it in a specific way to confinement ratio and volume fraction. Results from experiments and detailed simulations with different parameters (flexibility of different components in the suspension, viscosity ratio, confinement, among others) collapse onto the same curve. In simple shear flow, several regimes of segregation arise, depending on the value of a ``margination parameter'' M. Most importantly, there is a critical value of M below which a sharp ``drainage transition'' occurs: one component is completely depleted from the bulk flow to the vicinity of the walls. Direct simulations also exhibit this transition as the size or flexibility ratio of the components changes. Results are presented for both Couette and plane Poiseuille flow. Experiments performed in the laboratory of Wilbur Lam indicate the physiological and clinical importance of these observations.
Information geometry and the renormalization group
Host: E. Siggia
Microscopically diverse systems often yield to surprisingly simple effective theories. The renormalization group (RG) describes how system parameters change as the scale of observation grows and gives a precise explanation for this emergent simplicity in physics. We use information geometry to reformulate the RG as a statement about how the distinguishability of microscopic parameters depends on the scale of observation. We show that information about relevant parameters is preserved, with distances along relevant directions maintained under flow. By contrast, irrelevant parameters become less distinguishable under the flow, with distances along irrelevant directions contracting. We apply our tools to understand the emergence of the diffusion equation and more general statistical systems described by a free energy. This suggests a way to identify relevant directions for more general coarsening procedures.
Adaptation unifies emergent oscillations in quorum sensing populations
Host: E. Siggia
Adaptation is a ubiquitous trait of living organisms in dealing with the outside world. At the cellular level, high sensitivity over a broad range of environmental conditions is achieved via biomolecular circuits that operate out of thermal equilibrium. Here we report an unexpected link between sensory adaptation and auto-induced collective oscillations in dense cell populations. We uncover a frequency regime where adaptive cells amplify temporal variations of the extracellular signal. Deeply rooted in nonequilibrium thermodynamics, the design principle unites several known examples of dynamical quorum sensing, and provides a new perspective on regulatory mechanisms behind glycolytic oscillations in yeast cell suspensions.
Self-organization of curved and deforming active surfaces
Host: E. Siggia
Mechano-chemical processes in thin biological structures, such as the cellular cortex or epithelial sheets, play a key role during the morphogenesis of cells and tissues. Emergent dynamics in these processes can arise from a feedback loop in which active mechanical forces, by inducing material flows, indirectly affect their own chemical regulation. It has been demonstrated in simple, fixed geometries that this mechanism enables self-organized patterning, but the interplay of mechano-chemical processes with complex surface geometries and shape changes of the material remains to be explored. In this work, we employ the theory of active surfaces and develop analytical and numerical tools to study these materials in curved and dynamically evolving geometries. Additionally, diffusive and advective transport processes can redistribute molecules responsible for local stress generation within the surface, which resembles the interplay between active forces, the shape changes they imply and the effects this has on their regulation. Within this framework, cell polarization, contractile ring formation before cytokinesis, as well as pulsatile dynamics in active tubular structures can be understood as natural emergent phenomena. Our approach provides novel opportunities to explore different scenarios of mechano-chemical self-organization and can help understand the role of shape as an effective regulatory element in morphogenetic processes.
Building deep neural network models to understand biological vision
Host: C. Kirst
Recent advances in neural network modelling have enabled major strides in computer vision and other artificial intelligence applications. This brain-inspired technology provides the basis for tomorrow’s computational neuroscience [1]. Deep convolutional neural nets trained for visual object recognition have internal representational spaces remarkably similar to those of the human and monkey ventral visual pathway [2]. Functional imaging and invasive neuronal recording provide increasingly rich measurements of brain activity in humans and animals, but a challenge is to leverage such data to gain insight into the brain’s computational mechanisms [3, 4].
In my lab, we build neural network models of primate vision, inspired by biology and guided by engineering considerations [5]. We also develop statistical inference techniques that enable us to adjudicate between complex brain-computational models on the basis of brain and behavioral data [3, 4]. I will discuss recent work extending deep convolutional feedforward vision models by adding recurrent signal flow and stochasticity, two characteristics of biological neural networks. This improves inferential performance and enables neural networks to more accurately represent their own uncertainty.
[1] Kriegeskorte N (2015) Annu. Rev. Vis. Sci. 2015. 1:417-46.
[2] Khaligh-Razavi SM, N Kriegeskorte (2014) PLoS Computational Biology
[3] Diedrichsen J, Kriegeskorte N (2017) PLoS Computational Biology
[4] Kriegeskorte N, Diedrichsen (2016) J Phil. Trans. R. Soc. B.
[5] Spoerer CJ, McClure P, Kriegeskorte N (2017) Frontiers.
To move, or not to move, that is the evolutionary question: Evolution of growth and dispersal in bacterial populations
Host: T. Shendruk
I will present two projects demonstrating the interplay between growth and dispersal in evolutionary contexts. In the first one, I focus on the evolution of expanding populations, such as bacterial colonies or cancerous tumors. Population expansion is driven by growth and dispersal.
The evolutionary fate of a mutant also depends on both growth and dispersal. For instance, when could moving faster and growing slower be a favorable strategy? Starting with the Fisher-KPP equation, I come up with a quantitative rule of invasion, which depends on growth and dispersal rates of the ancestor and the mutant. Theoretical findings are supported by experiments with bacterial swarming colonies, as well as data from ecology literature.
In the second project, I address the question of design optimization of a biochemical network (the C-di-GMP network in bacteria) that integrates signals from the environment and regulates genes for growth and dispersal. This network needs to be trained to provide the optimal outcome when the environment alternatively favors growth or dispersal. I show that natural selection in this network architecture is mathematically equivalent to training a machine learning model to solve a classification problem.
Developmental biology is at a special point in its history: 1) Model organisms have been extensively characterized, 2) the molecular parts list has largely been enumerated, and 3) live-imaging and sequencing tools have matured. What remains is to understand how the system works. I will present work in progress in collaboration with Richard Carthew that focuses on a diverse set of phenomena made manifest during Drosophila eye development. Incomplete results related to collective cell-fate determination, collective polarization, and mechanics will be presented. An emphasis will be placed on the importance of taking a dynamical view of development, and mathematical modeling as a way to synthesize disparate observations, guide experimentation, and make predictions.
Evolution and motor control in bat flight - A wing and a prayer?
Host: T. Shendruk
Bat wings evolved from grasping, manipulating mammalian hands, and this origin influences the biomechanics of flight in bats in comparison to flight in birds and insects. Therefore, an evolutionary perspective is critical to advancing the comparative biology of flight, and helps distinguish those aspects of flight that are shared in all flying animals and those features that are unique to bats. Low weight, particularly in the wings, is important for all flying animals, but selection for reduced wing mass in bats must interact with aspects of neural control in the most morphologically complex of animal wings. In addition, the nature of wing skin as a complex functional material and the capacity to modulate wing mechanical properties during flight by an unusual group of muscles found only in bats proves critical to bat flight performance. Improved understanding of the functional architecture of bat wings not only provides insight into steady-state flight behaviors, but also holds promise for solving problems concerning bats’ abilities to recover from perturbations, fly effectively even following wing damage or injury, etc. This approach requires sophisticated bioengineering techniques such as particle image velocimetry, multi-camera high speed videography, and dynamic modeling, but also low-tech methods including polarized light photography, histology, and anatomical description.
Wisdom of hives and mounds: collective problem solving by super-organisms
Host: T. Shendruk
Social insects are capable of solving complex physiological problems using collective strategies. I will discuss our work on some of these problems that include the physiology and morphogenesis of termite mounds, and active mechanisms for ventilation, mechanical adaptation and thermoregulation in bee aggregates.
The Fast and the Furious: Mechanics and dynamics of rapid cell motility
Host: T. Shendruk
Directed crawling motility of animal cell types ranging from neurons to macrophages requires the coordinated force-generating activity of multiple mechanical elements. Much molecular detail is now known about the constituents of some mechanical submachines such as the polymerizing actin network and the adhesion complexes, but it is not yet clear how these elements all work together to generate coherent, directed motion at the level of the whole cell. In order to understand cellular mechanisms of large-scale coordination, our work focuses on two extremely fast-moving cell types, the fish epidermal basal keratocyte responsible for the rapid closure of wounds in fish skin, and the human neutrophil that hunts down and kills microbial invaders. Despite their very different biological roles and apparent behaviors, these cell share fundamental biophysical mechanisms of self-organization and movement coordination.
Dryland landscapes show a variety of vegetation pattern-formation phenomena, two striking examples of which are banded vegetation on hill slopes and nearly hexagonal patterns of bare-soil gaps in grasslands (“fairy circles”). Vegetation pattern formation is a population-level mechanism to cope with water stress, that is coupled to other response mechanisms operating at lower and higher organization levels, such as phenotypic changes at the organism level and biodiversity changes at the community level. Uncovering the roles that vegetation pattern formation plays in the functioning of dryland ecosystems is a challenging problem of particular significance in the current era of climate change and massive human intervention in natural ecosystems. In this talk I will present a platform of mathematical models for dryland ecosystems and use it to study (i) mechanisms of vegetation pattern formation, (ii) the variety of extended and localized patterns that can appear along the rainfall gradient, (iii) the impact of pattern formation on ecosystem response to droughts, and (iv) forms of high-integrity human intervention that do not impair ecosystem function. Universal aspects that might be applicable to other living systems will be emphasized.
>
The mathematics of biomedical and biophysical imaging
Hosts: M. Magnasco and S. Strickland
Over the last few decades, new mathematical techniques have played an important role in a variety of biomedical and biophysical imaging modalities. Following a brief survey of the field, we will focus on recent progress in nonlinear optimization that is bringing large-scale problems in acoustic scattering and cryo-electron microscopy within practical reach.
Tracing lineage and cell differentiation at single cell resolution
Host: T. Shendruk
I will present our efforts to map cellular differentiation hierarchies in zebrafish and xenopus embryos, with further examples in mammalian tissues. A few years ago we (and others) developed droplet-microfluidic technology for single cell RNA-Seq (inDrop), opening up the possibility to infer cell states in an unbiased manner. We recently carried out single-cell RNA sequencing of >200,000 cells from time series of the first 24 hours of life, in two vertebrate species: zebrafish (Danio rerio) and the frog (Xenopus tropicalis). We reconstructed the first tens of cell fate choices in the embryo from axis patterning, germ layer formation, and early organogenesis, and compared the two organisms in their differentiation dynamics. We tested how clonally-related cells traverse these fate choices by developing a transposon-based barcoding approach (“TracerSeq”) for reconstructing single-cell lineage histories. Through different examples, I will try to show some of the challenges and opportunities of single cell RNA-Seq approaches: finding and validating new cell types, identifying new growth factor regulators of fate choice, clarifying points of fate commitment of tissues, but also showing where single cell RNA-Seq data can be misleading.
