Oral Presentation Abstracts
-
Authors: Chalyshkan S., Biswas T., James B., Chen Y., Förster D., Stockert L., Bispo P., Kubo F., Baier H., Fitzgerald J.E.
Abstract: With the growing use of large-scale activity recordings and advances in anatomical circuit reconstruction, researchers can develop neural network models linking synaptic connectivity to neuronal activity. However, determining how well different model classes capture observed structure-function relationships remains a major challenge. Rather than fitting a single model, here we assess the fidelity of threshold linear recurrent neural network models by testing predictions from an ensemble modeling approach that identifies features consistent across all admissible models1,2. We specifically apply this framework to optic flow processing in the larval zebrafish pretectum, leveraging a function-linked (FuL) connectomics dataset with synaptic resolution alongside functional recordings3,4. We first validated the framework by representing the pretectal circuit at the level of major functional cell types and evaluating predictions of neural activity and synaptic connectivity. Robust structure-function relationships emerge at the level of neural activity. However, connectivity predictions remain sensitive to model assumptions and are yet to be conclusively resolved. We therefore describe ongoing work to further characterize the pretectum through additional data analysis of the FuL connectomics dataset and improved cell type definitions. We are also developing novel approaches to quantify uncertainty arising from population coarse graining to seamlessly make single neuron predictions from population level modeling.
-
Authors: Christos Zangos, Danish Ebadulla, Thomas C. Sprague, Ambuj K. Singh
Abstract: Reconstructing natural images from fMRI is a powerful way to characterize how the brain represents visual information, but state-of-the-art methods require tens of hours of scanning per participant, which is impractical for most cognitive neuroscience laboratories. We introduce Adapter Alignment (AA), a training procedure that calibrates a new participant to a shared representational space learned from a densely sampled reference participant. Lightweight subject-specific adapters are first explicitly aligned to the reference on a small set of shared calibration images, and the full reconstruction pipeline is then fine-tuned while this alignment is preserved. We pair AA with a coverage-based algorithm that selects calibration images spanning the principal dimensions of the shared space. With one hour of data, AA improves reconstruction metrics by up to 8 percentage points and pre-diffusion image retrieval by up to 9.4 points; combined with image selection, it surpasses a one-hour baseline using only 24 minutes of subject-specific data. AA transfers across architectures and datasets, including adapting 3T participants to a 7T reference, bringing image reconstruction within reach of studies of how attention and working memory shape visual representations.
-
Authors: Cyrille Mascart, Khue Tran, Khristina Samoilova, Liam Thomas Storan, Tingkai Liu, Alexei Koulakov
Abstract: Understanding the structure of perceptual odor space remains a central challenge in olfaction. Unlike vision or audition, where perceptual spaces can be organized along a small number of perceptual dimensions, the dimensionality, geometry, and organizing principles of human olfactory perception remain poorly understood. Recent advances in deep learning have enabled prediction of odor percepts from molecular structure, but most existing approaches represent perception using fixed sets of semantic labels. Such representations may capture prominent perceptual categories while discarding information contained in the richer structure of natural odor descriptions, potentially limiting our ability to characterize perceptual relationships among odors.
-
Authors: Danna Pinto, Gal Vishne, Robert Knight, Leon Deouell
Abstract: Previous human intracranial studies have shown that visual category information remains sustained and temporally stable in high-frequency activity (HFA) throughout stimulus presentation in occipitotemporal cortex, even as response magnitude declines [1-3]. However, it remains unclear whether this sustained representation is specific to HFA or distributed across the neural frequency spectrum. Here, we asked whether sustained category information extends across multiple frequency bands, whether representations generalize between bands over time, and whether these cross-band effects reflect band-specific activity or broader aperiodic spectral changes.
-
Authors: Derek Nguyen, Yi Zhou
Abstract: Understanding how the brain represents communication sounds is a central challenge in auditory neuroscience. Like human speech, marmoset conspecific calls contain complex spectral features, including global harmonic structure and local frequency modulation, that carry information about caller identity and voice. It remains unclear how cortical neurons encode these natural features and whether the same neural code represents vocalizations and non-vocal sounds. We synthesized marmoset calls using a virtual marmoset vocalization model [1] and systematically manipulated harmonic structure and frequency modulation to create a stimulus continuum morphing from call-like to noise-like sounds. Single-unit activity was recorded from auditory cortex of two adult marmosets using 128-channel silicon probes. We quantified neural representations using the Victor–Purpura spike-distance metric [2] and hyperbolic multidimensional scaling (HMDS) [3], allowing us to examine how firing rate and precise spike timing shape representational geometry. Three observations emerged. First, spike timing carried substantial information about both calllike and, surprisingly, noise-like sounds. Second, neurons differed markedly in their reliance on rate versus timing. Third, neural representations shifted systematically from call-like toward noise-like response patterns across the stimulus continuum. These findings suggest that cortical encoding of communication sounds depends critically on temporal spike patterns, not simply response magnitude or firing rate.
-
Authors: Eirini Schoinas, Christos Zangos, Thomas Christopher Sprague, Ambuj Singh
Abstract: Computational models may share the brain’s organization of visual stimuli without following the same sequence of transformations. We introduce an alignment-free Riemannian framework that separately compares representational geometry and processing order across human EEG/MEG responses, an imagefed vision model (DINOv2), and a label-fed language model (Llama). Neighborhood covariance trajectories describe each stimulus’s local representational structure, and pairwise velocity angles enable comparison without aligning neural and model coordinates. Across ten EEG and four MEG participants, both models shared positive angular geometry with neural responses. DINOv2 showed higher mean geometric concordance, whereas rate-free ordering tests yielded more consistent evidence for Llama. Category and stimulus-neighborhood controls qualified the interpretation of this agreement. These findings distinguish similarity in neural representational geometry from correspondence in processing order.
-
Authors: Frances S. Chance, Suma George Cardwell
Abstract: The survival of an animal can depend on its ability to perform fast calculations with limited energy resources, for example as part of an essential behavior such as predation or escape. What biophysical mechanisms support these fast and energy-efficient computations in biological nervous systems? Our approach uses neuromorphic emulation to identify key computations performed by biological systems and demonstrate the advantages of incorporating neural-inspired computational primitives into engineered systems.
We have specifically focused on shunting inhibition, a membrane conductance with reversal potential close to the resting membrane potential of a neuron, as a proposed neuromorphic primitive. Shunting inhibition is a putative mechanism by which a biological neuron multiplies two time-varying signals, for example in multiplicative gain modulation [8]. We have developed a neuromorphic emulation [9, 10] of shunting inhibition described in Drosophila T4 neuron dendrites [7]. In our model, subthreshold analog field effect transistors (FETs) approximate conductance changes driven by synaptic input to a biological neuron. Shunting inhibition may therefore be controlled via the gate terminal (see Figure 1) of the appropriate transistor. When visual inputs are modeled such that their receptive fields are in a similar spatial configuration to those to the T4 neuron, the direction-selectivity of the neuromorphic neuron is qualitatively similar to that of the T4 neuron [11].
-
Authors: G. William Chapman, Andrew S. Alexander, Frances S. Chance
Abstract: Animals intercept moving targets by continuously estimating target position, target motion, and how self-motion changes these relationships. Here we demonstrate that the neural substrate of this latent state depends on the available sensorimotor variables: dendritic state is critical when target location must be maintained internally, whereas hierarchy and temporal filtering can suffice when egocentric vision continually resupplies target information. Building on experimental and modeling work in spatial navigation and target pursuit, we train closed-loop networks under two regimes. In target-motion, the agent receives nonegocentric target-motion variables and must construct egocentric targets for action. In egocentric-vision, the agent receives visual input and must construct latent target and self-position representations. We compare single-region and hierarchical networks with tanh, leaky-integrator, LIF, and dendrite-augmented units, and analyze allocentric, egocentric-boundary, and egocentric-target tuning. Hierarchy promotes compact spatial codes when intermediate variables must be built, while leaky and dendritic dynamics stabilize representations across time. Dendrite-augmented units outperform baseline LIF and tanh models in target-motion. A local dendritic self-supervised learning rule, comparing feedforward activity with top-down dendritic predictions, recovers similar latent variables. These convergent mechanisms yield testable predictions for experimental-computational studies of pursuit, navigation, and dendritic contributions to spatial codes.
-
Authors: Greg D. Field, Kiersten Ruda, Andra M. Rudzite, David St-Amand, Na Young Jun, John Pearson
Abstract: Efficient coding theory has long provided an influential account of why retinal ganglion cells (RGCs) develop antagonistic center-surround receptive fields (RFs): decorrelating spatially correlated visual input improves the efficiency with which the optic nerve’s limited bandwidth is used. Classic treatments of this idea [1–4] predict that as signal-to-noise ratio (SNR) declines, for example moving from cone- to rod-mediated vision, RF surrounds should weaken and eventually disappear, leaving a purely low-pass, center-only filter that pools more photons to combat noise. We tested this prediction directly using large-scale multi-electrode array recordings of six simultaneously measured RGC types in rat retina across a million-fold range of light levels (0.1–10,000 Rh*/rod/s). Contrary to the classical prediction, five of six RGC types retained robust, band-pass center-surround RFs even at the lowest light levels tested.
-
Authors: Hannah Choi, Jun Yang, Kseniia Shilova, Drew Lawless, Cem Sevinc, Chris C. Rodgers
Abstract: To understand the world around us, we use body movement to actively probe for information. For instance, you might grasp each object in your bag in turn, until you recognize your keys by their texture and shape. To do this, the brain must synthesize sensory and motor signals to generate an internal model of the world and to guide adaptive action. However, the neural circuits for sensorimotor integration remain poorly understood. In this project, we are investigating how neural circuits in the sensory and motor cortex implement the computations of active sensing by combining data-driven modeling with experiments in mouse whisker system.
-
Authors: Ivy Wang, Christos Zangos, Thomas C. Sprague, Ambuj K. Singh
Abstract: Linear encoding reveals whether caption-derived large language model (LLM) features predict neural responses, but not whether brain and model representations share local geometry. Using 7T fMRI responses from all eight Natural Scenes Dataset subjects and Llama-3-8B representations of captions describing the viewed COCO images, we compare voxel-wise encoding with the alignment of local variation subspaces estimated within semantic neighborhoods. Ridge regression yields positive held-out prediction accuracy across cortical regions, with higher accuracy in body-selective than early visual cortex and little variation across Llama layers. Within each system, subspace orientation changes systematically along the cortical ROI ordering and with Llama layer separation. Across systems, alignment remains near the isotropic random-subspace expectation, but is consistently stronger in EBA and FBA than in early visual cortex. Unlike Llama, CLIP text-encoder representations show increasing encoding accuracy and stronger alignment with depth. These findings dissociate predictive from geometric correspondence: cortical responses can be linearly predicted despite weak alignment of the dominant local variation subspaces.
-
Authors: Jae Hyung Woo, Michael C. Wang, Peter Duggins, Ramon Bartolo, Daeyeol Lee, Bruno B. Averbeck, Alireza Soltani
Abstract: Flexible behavior relies on the brain’s ability to use reward feedback to form and update expectations that guide future choices. Beyond updating expectations, reward also shapes neural processing to prioritize behaviorally relevant information. But how are expectations updated and relevant information prioritized when the underlying causes of reward are uncertain? To address these, we recorded neural activity in the prefrontal cortex of monkeys performing a learning task in which reward outcomes depended on uncertain associations with either chosen stimuli or actions. We found that monkeys updated both associations in parallel, and accumulated differences in reward prediction errors to estimate the reliability of strategies. Prefrontal neurons encoded this reliability, which influenced choice by arbitrating between learning systems. This was driven by both trial-by-trial reward feedback and reliability belief that dynamically aligned value representations with the choice axis when relevant and rotated them away when irrelevant. Using neural network models, we reproduced the behavioral and neural data, showing that these mechanisms are feasible and robust, and that representational geometry plays a causal role in arbitration. Our results show that reward feedback reshapes how the brain represents information over both short and long timescales, helping the most appropriate learning strategy guide behavior.
-
Authors: Hohyun Cho, Carl S. Townsend, Anas U. Khan, George R. Tamula III, Muayad A. Alzuabi, Benjamin C. Cox, Peter Brunner, Robert T. Knight, Hal Blumenfeld, Jasmine A. Thum
Abstract: Background. Impaired consciousness (IC) during focal seizures is a major source of disability in epilepsy. The network inhibition hypothesis proposes that focal seizures can disrupt midline subcortical arousal systems, leading to bilateral frontoparietal cortical suppression and IC [1]. However, the roles of individual basal ganglia nodes and the dynamics that mark the transition to IC during focal seizures remain poorly defined. We therefore tested whether theta-band activity and network synchronization within this mesocircuit distinguish seizures with IC from seizures with preserved consciousness (PC).
-
Authors: Md Abu Bakr Siddique, Jakub Orłowski, Yan Zhang, and Hongyu An
Abstract: Parkinson’s disease (PD) is a progressive neurodegenerative disorder involving loss of dopaminergic neurons and impaired movement, including tremor, rigidity, and bradykinesia. Medications such as levodopa are the primary treatment, but motor fluctuations and dyskinesias can emerge over time. For selected patients with disabling symptoms despite medication, deep brain stimulation (DBS) provides an additional therapy by electrically modulating basal-ganglia circuits, often in the subthalamic nucleus (STN) [5].
Conventional DBS typically applies continuous pulses at preprogrammed settings even when the pathological neural state changes. This open-loop operation can increase battery consumption and stimulation-related side effects. Adaptive DBS (aDBS) instead uses neural feedback to adjust stimulation as needed, commonly tracking abnormal beta-band (13–30 Hz) activity in STN local field potentials (STN-LFPs) [4,5]. A key challenge is implementing responsive, reliable biomarker-driven control within implantable power and size limits: most proposed approaches remain software-centric, with limited transistor-level realization.
-
Authors: Michael D. Vivian, Michael Finch, Aaron D. Milstein, Christine Grienberger
Abstract: How do neural circuits rapidly incorporate new information while limiting further plasticity as useful representations become established? In hippocampal area CA1, behavioral timescale synaptic plasticity (BTSP) can generate a place field after a single experience through synaptic weight changes triggered by dendritic plateau potentials (“plateaus”)1,2. Previous recordings during spatial learning in head-fixed mice navigating a linear treadmill for water reward revealed corresponding patterns of entorhinal cortex layer 3 (EC3) activity and CA1 overrepresentation near reward in a cue-rich environment and near a reward predictive visual cue positioned 50 cm before reward in an otherwise cue-depleted “sparse” environment. Together with EC3’s role in promoting plateau initiation, these results support the idea that EC3 provides an environment-specific target signal that drives BTSP and thereby CA1 representation formation during spatial learning. The observed place field distributions and declining formation rates provide spatial and temporal targets for testing how relatively stable EC3 input interacts with dendritic inhibition to shape spatial learning.
-
Authors: Mingxiao Wei, Ryan Golden, Wojciech Goch, Pavel Sanda, Giri P. Krishnan, Oxana Eschenko, Maxim Bazhenov
Abstract: Sleep plays a key role in brain function, including memory consolidation yet the mechanisms by which neuromodulatory systems regulate memory-related brain activity during sleep remain poorly understood. During slow-wave sleep (SWS), cortical Up states, the active phase of the sleep slow oscillation (SO)3, provide a critical temporal window for coordinating thalamocortical spindles, hippocampal ripples, and hippocampal-cortical replay to support memory consolidation4. The duration of this window has been shown to play an important role in the outcome of memory consolidation, with shorter Up states biased toward novel context reactivation. While the locus coeruleus-norepinephrine (LC-NE) system is traditionally associated with wakefulness and arousal, recent evidence indicates that LC activity is dynamically structured during SWS and may regulate this Up-state-centered coupling window. Here, we tested whether LC-NE modulation of cortical excitability controls Up-state duration in a large-scale biophysical model of the cortico-thalamo-hippocampal loop9 (Figure 1A), build on rodent data. During SWS, LC neurons exhibit rhythmic population bursting, which drives corresponding infraslow (~0.05 Hz) fluctuations in extracellular NE levels10. The inter-burst interval distribution from rodent SWS is shown in Figure 1B, and we modeled LC bursts according to this empirical distribution (Figure 1C, top). Because burst-evoked NE peaks can recruit low-affinity β-adrenergic receptors, we modeled the cortical βadrenergic effect of LC burst-evoked NE as suppression of the effective calcium-activated potassium after hyperpolarization conductance (G_KCa) in cortical pyramidal neurons (Figure 1C, bottom) and examined how this modulation alters cortical Up states (Figure 1D,E). Stronger LC-NE-mediated suppression of G_KCa prolonged Up states, whereas reducing this suppression preserved stronger KCamediated hyperpolarizing currents and shortened Up states (Figure 1F). In a parallel experimental analysis in rodents, cortical gamma events, defined by elevated 30-40 Hz EEG power, were temporally linked to LC bursts. Because these gamma events mark a subset of multiunit-activity-defined cortical Up states, gamma event duration provided an experimental readout of Up-state-associated cortical activation. Longer delays from LC burst to gamma onset predicted shorter gamma event duration (Figure 1G), consistent with a transient effect of LC-NE signaling that prolongs cortical Up states following LC bursts. In a parallel study, using computational models and data from urethane-anesthetized animals, we found that longer Up states promote reactivation of both familiar (cortical) and novel (hippocampus-dependent) memories, whereas shorter Up states are dominated by replay of novel (recent) memory traces5,6. Together, the model and data support a mechanism in which precisely timed LC-NE signaling shapes SWS network dynamics, potentially influencing the outcome of memory consolidation.
-
Authors: Oliver Gambrell, Eckhard Friauf, Abhyudai Singh
Abstract: Neurotransmission is critically influenced by the statistics of quantal content (QC), i.e., the number of synaptic vesicles (SV) fusing at the active zone of a presynaptic axon terminal upon the arrival of an action potential (AP). In the mammalian auditory brainstem, inhibitory glycinergic MNTB-LSO synapses play an important role in sound localization [1,2]. Previous studies at these synapses have shown that a single readily-releasable pool of SVs can describe the steady-state statistics of QC [3]. Here we investigate the transient statistics of QC at MNTB-LSO synapses by performing voltage-clamp electrophysiological recordings at the same single cell in response to 20 repeated stimuli (100 stimuli at 100 Hz for 1 sec). Our data shows that the repeat-to-repeat variability in QC, as quantified by the Fano factor (FF; variance over mean), increases for the first ten stimuli and then plateaus at a fixed value. This observed increase in FF was much smaller than predicted by the single-pool model. Model fitting and comparison using the Akaike Information Criterion [4] revealed two different readily-releasable pools [5] of SVs with different transient probabilities of SV docking at empty sites (i.e., docking probabilities). Our approach shows how the stochastic dynamics of QC can be leveraged to uncover mechanisms regulating neurotransmission.
-
Author: Paul Nuyujukian
Abstract: Latent-variable methods have transformed systems neuroscience by exposing low-dimensional structure in population activity, yet the biological meaning and stability of these dimensions remain unclear. This talk examines how latent dimensions relate to behavior, network dynamics, and physical circuitry. In motor cortex, supervised two-dimensional latents capture only 3–7% of neural variance yet decode reaching better than common alternatives and remain stably related to behavior over years, despite drift in neuron-to-latent encoding (Abdulla et al., Nature Communications, accepted). Stiefel Manifold Dynamical Systems explicitly model such nonstationarity by allowing orthonormal emission matrices to evolve while sharing latent dynamics, revealing slower drift in behaviorally important dimensions (Lee et al., Neural Computation, accepted). On faster timescales, dominant neural subspaces rotate and state eigenvectors exchange rank as neuronal contributions fluctuate; subspace alignment rescues decoding, while network models implicate excitatory coupling and spike-rate adaptation in this continuous reparameterization (Clarke et al., revision submitted). Finally, matched functional and connectomic datasets in mouse visual cortex and C. elegans show that latent activity patterns predict direct synaptic connectivity above chance (Nuyujukian, in preparation). Together, these findings recast neural manifolds as behaviorally stable but dynamically reconfigured objects, constrained by task complexity and grounded partly in physical circuits.
-
Authors: Sai Sun, Tao Xie, Shuo Wang, Peter Brunner, Jon T. Willie
Abstract: Introduction. Social-affective decision-making, including the categorization of subtle facial expressions under ambiguity, is fundamental to human social functioning. These decisions can be understood within computational frameworks in which noisy evidence is gradually accumulated toward a decision threshold—a process known as evidence accumulation (EA)1. Although the amygdala is strongly implicated in social-affective processing, whether it selectively implements and causally modulates EA remains unknown. We hypothesized that the human amygdala selectively implements emotional EA, with neural signatures organized in a hemisphere- and nucleus-specific manner, and that amygdala perturbation causally modulates this computation. We tested this hypothesis in neurosurgical patients by combining intracranial recordings, computational modeling, and causal perturbations.
-
Authors: Sanjoy Dasgupta, Yaron Meirovitch, Xingyu Zheng, Inle Bush, Jeff W. Lichtman, Saket Navlakha
Abstract: Finding a maximum-weight bipartite matching is a fundamental problem in combinatorial optimization, with applications from assigning reviewers to papers to matching hospitals to medical students. Existing algorithms rely on centralized computation, shared memory, or unique node identifiers. We discovered that during the development of the neuromuscular circuit, a bipartite matching is formed using a much simpler, distributed algorithm. In the adult, N motor neurons and M ≥ N muscle fibers form a “1-to-many” matching: every fiber is innervated by exactly one neuron, while each neuron is typically connected to more than one fiber. At birth, however, connectivity is much denser (e.g., about 50% of all possible connections at birth in the mouse interscutularis muscle [3]). The matching architecture emerges through activity-dependent competition: neurons innervating the same fiber compete for synaptic territory, and resources withdrawn from losing synapses are redistributed to the neuron’s remaining synapses [2]. We formalized this pruning process as a distributed matching algorithm and showed that it rivals centralized matching algorithms and recapitulates features of neuromuscular development.
-
Authors: Tal Dalal, Or Shachar, Yoav Nadu, Shany Nivinsky-Margalit, Ido Paz, Noham Wolpe, Ran Darshan, Arseny Finkelstein
Abstract: Goal-directed behavior require integration of effort costs with expected rewards to decide whether to persist in goal pursuit or disengage. These effort-reward contingencies are rarely signaled explicitly and can persist over long timescales, requiring inference from experience and memory maintenance of their estimated value. Yet, the circuit mechanisms underlying effort-reward integration to shape goal value and motivational states remain unknown. Here, we developed a novel behavioral paradigm requiring no training, in which mice dynamically infer changing latent effort-reward states and flexibly switch between engagement, disengagement and re-engagement. High-density recordings in the anterior lateral motor cortex (ALM) and orbitofrontal cortex (OFC) revealed dynamic representations of latent state value. Population analyses identified persistent activity graded by the subjective value of the current effort-reward state. During state transitions, activity tracked the gradual updating of motivational state before converging to a new value set-point. Unilateral ALM perturbation induced rapid recovery of activity to its value set-point, consistent with continuous attractor dynamics. In contrast, bilateral perturbation abolished setpoint recovery and rescaled behavioral performance. Together, these findings identify persistent, value-graded dynamics in frontal cortex that underlie the maintenance of motivational state.
-
Authors: Teng Fei, Srinivas Ravishankar, Zhining Chen, Abhinav Uppal, Ian Jackson, Garrison W. Cottrell, Virginia R. de Sa
Abstract: What visual information can be read out from EEG when a person is looking at time-locked images? EEG is known to be a noisy and spatially blurred signal of the underlying brain activity; however, averaging responses has long been used to increase the signal-to-noise ratio of these recordings. Recently, we and others have been exploring which aspects of visual inputs can be decoded from EEG (after averaging many time-locked signals) [1]. In this work, we explore different latent spaces as useful filters for extracting information from the EEG (using the public THINGS-EEG2 dataset [2]). By learning encoding mappings from different latent spaces to the EEG, we can explore when and where different features are encoded in the EEG. By learning decoding mappings from EEG to different latent spaces, we can reconstruct predictions of the observed images. Linear decoding mappings provide high-quality reconstructions while being easy to analyze. Linear encoding maps can easily be used to visualize presumptive directions in the latent space that give increased positive-going or negative-going EEG voltage at any electrode and time bin.
-
Author: Vaibhav A. Diwadkar
Abstract: Brains rely on their capacity to dynamically change network configurations as they adaptively optimize responses to unpredictable inputs (Muzik & Diwadkar, 2023; Sprevak & Smith, 2023). Psychopathologies like schizophrenia and obsessive compulsive disorder (OCD) alter this adaptive capacity and may well be characterized by altered brain network dynamics (Uhlhaas, 2013). Here we motivate two complementary approaches, to characterize aberrant task-evoked network dynamics from fMRI data. Centrality Dynamics (CD) (Bhatt et al., 2025) is based on capturing changes in the centrality of brain nodes over the course of task dynamics. To estimate CD, first connectivity matrices (based on zero lag functional connectivity) were estimated from fMRI time series data acquired during learning (Nanaaware et al., 2025), in each of 246 brain regions (nodes)(Fan et al., 2016) and in 280 partially overlapping temporal windows (where each window represents a frame of evoked brain activity). Next, we estimated every node’s betweenness centrality (BC)(Meram et al., 2023) in each of these 280 temporal windows. From the estimated BC, time series were formed, henceforth tBC (246 nodes, 280 points each) where each tBC captures a node’s CD across the task. Across controls and patients, these tBC were clustered to recover sub-networks homogenous centrality dynamics. We show that these sub-networks in schizophrenia show lower intra-cluster coherence of CD, were formed by nodes with greater functional heterogeneity, and show greater stochasticity of CD.
-
Authors: Vinay S. Raghavan, Jens Madsen, Maximilian Nentwich, Marcin Leszczynski, Arnaud Falchier, Stephan Bickel, Brian E. Russ, Lucas C. Parra
Abstract: Primate vision is an active process. We sample the world through a series of saccadic eye movements and brief fixations.1 While the integration of low-level visual features across saccades is well-documented,2 the mechanism by which the brain maintains semantic continuity across multiple ”snapshots” remains a fundamental mystery. Inspired by recent breakthroughs in self-supervised learning,3 we hypothesized that the primate brain employs a predictive mechanism to bridge successive fixations. In this framework, the brain predicts the semantic content of an upcoming fixation; a ”semantic novelty” signal is then generated to update internal representations. Specifically, this novelty signal is computed only at the semantic level and not at lower levels of representation. Here, we define novelty as the distance between embeddings of successive foveal fixation patches and measure whether novelty computed at low-, mid-, and high-levels of representation modulates fixation-locked activity.
-
Authors: Xiaoxiao Lin, Miranda De Leon, Luhan Wang, Akil Wijeratne, Jarod Bussey, Jingheng Lu, Andrew S. Alexander
Abstract: Cortical circuits support predictive pursuit of moving targets. Xiaoxiao Lin, Miranda De Leon, Luhan Wang, Akil Wijeratne, Jarod Bussey, Jingheng Lu, Andrew Alexander. Department of Psychological and Brain Science, University of California, Santa Barbara Successful navigation in dynamic environments requires animals to predict the future locations of moving targets based on internal representations of space and self-motion. Rather than continuously tracking a target’s current position, animals can use predictive strategies to intercept targets more efficiently. However, the neural mechanisms that support predictive target interception during naturalistic navigation remain poorly understood. To address this question, we trained mice in a reward-associated visual target pursuit task in which a moving target followed either pseudorandom (unpredictable) or structured (predictable) trajectories. Mouse position and target motion were tracked in real time using DeepLabCut, and successful interceptions triggered water reward delivery. Behaviorally, mice initially pursued targets reactively based on immediate visual feedback. With training on structured trajectories, animals progressively adopted predictive strategies, initiating movement toward anticipated future target locations rather than the target’s current position. These behavioral changes suggest the emergence of an internal model of target dynamics relative to the animal’s own movement. To investigate the neural basis of this transition, we performed high-density electrophysiological recordings in the cingulate cortices and hippocampal formation, regions implicated in spatial representation, navigation, and goal-directed action planning. Using single-neuron and population-level analyses, we examined how neural activity encoded spatial variables and whether population dynamics reflected predictive representations of future target locations. Together, these findings provide insight into how cortical circuits support the formation of predictive internal models that guide flexible, goal-directed pursuit behavior.
-
Authors: Zach Saccomano, Sam McKenzie, Daniel English, Horacio G. Rotstein, Asohan Amarasingham
Abstract: Individual synapses can leave a striking signature in the fine-timescale structure of presynaptic and postsynaptic spike trains, illustrated by the “monosynaptic” peak in the spike cross-correlogram (CCG): timelagged synchrony concentrated with millisecond-scale precision on lags that match synaptic conduction delays. Connecting this signature to synaptic properties will be a routine challenge for large-scale spike train recordings for the foreseeable future. To isolate it, we adopt a causal modeling perspective, explicitly distinguishing a spontaneously recorded spike train pair from its counterfactual counterpart had the monosynaptic connection been silenced. This perspective shows that the monosynaptic information contained in lagged synchrony statistics is fundamentally entangled with firing rates, presynaptic bursting, and nonstationary background dynamics; CCG-based methods cannot, in general, identify causal synaptic effects. Instead, we formulate explicit models of a synapse’s causal effect on the fine-timescale structure of postsynaptic spike trains, assuming only that non-monosynaptic background influences operate on coarser timescales, while taking care to minimize all other assumptions. Under these assumptions we derive unbiased estimators, exact confidence intervals, and fast inference algorithms, together with gradient techniques for fitting model hyperparameters, and test the framework against a wide spectrum of neural dynamical systems models. Inference is robust, and remains accurate when the background processes are confounded or nonstationary on timescales as fast as tens of milliseconds, which is the major challenge of in vivo analysis. We will show in simulations that this approach achieves superior performance to the state of the art.
-
Author: Zhongming Liu
Abstract: The brain’s “resting” state is an “active” state of interoception The brain is never isolated from the body. Even in the absence of external stimuli or an explicit behavioral task, it continuously senses and regulates the heart, lungs, gastrointestinal tract, and other internal organs through interoception. Nevertheless, resting-state brain activity and its temporal correlations are often interpreted in terms of intrinsic neural dynamics and functional connectivity. Many functional magnetic resonance imaging (fMRI) studies focus on this interpretation to map the human brain connectome. I challenge this interpretation and advocate for a different view. That is, the “resting” state is not intrinsic to the brain, but an active state of “interoception”. Supporting this view, I present experimental evidence from animal studies and computational models based on predictive coding.
-
Authors: Xiaohan Zhang, Alex Doran, Zhe Sage Chen
Abstract: In recent years, digital twin brain (DTB) has emerged as an important research area that leverage the exponential growth of knowledge in brain connectomes, large-scale neural recordings, deep learning, and neurostimulation [1]. By modeling the whole brain or localized circuits in silico, the digital twin represents a biological replica, which may provide mechanistic insights into functions of healthy and diseased brains as well as personalized treatment such as deep brain stimulation (DBS). Motivated from the mouse behavioral and electrophysiological data [2,3], we have previously developed a biologically constrained, cell-type specific neural network as a digital twin for the mouse prefrontal cortex and mediodorsal (MD) thalamus in cognitive control and flexible decision making [4]. We showed that the addition of a feedforward MD structure (“controller”) to the recurrent PFC increases robustness to low cueing signal-to-noise ratio (SNR), enhances working memory, and enables rapid context switching. The model replicates key neurophysiological findings in task-performing mice, revealing computational mechanisms of MD in regulating cue uncertainty and context switching to enable cognitive flexibility [4]. We recently extended this model via adding two important features: (i) incorporating multiplicative gating mechanism [5]; (ii) generalizing rate-based to spiking-based neural networks. First, the spiking neural network model replicates the key findings of the prior rate-based neural network model. Next, we show that emergent rhythmic couplings emerge in a large-scale (N=~1000-2000 spiking neurons) task-performing PFC-MD network, where the LFP proxy may serve as a robust readout of interareal communication. Furthermore, different thalamic gating mechanisms (additive alone vs additive + multiplicative) yield qualitatively and quantitatively different thalamocortical LFP couplings in both amplitude and phase. Finally, from a control perspective, thalamic stimulation provides a means to modulate cortical dynamics in cognitive control and enhance behavioral flexibility of disease phenotypes. Our DTB model provides a platform for in silico MD stimulation for testing different open- or closed-loop thalamic stimulation strategies and providing experimentally testable hypotheses in animal models.
Acknowledgment: This work was supported by the NIH CRCNS grant R01-MH139352.
Poster Presentation Abstracts
-
Authors: Katherine Doxey, Derek Nguyen, Yi Zhou, Sharon Crook
Abstract: Advances in transcriptomic, morphological, and electrophysiological techniques have substantially improved the characterization of cortical neuron types. However, it is challenging to identify neuron type based on extracellular action potentials (EAPs) recorded from single units without ground-truth labels. Existing classification methods that rely on extracting features and applying machine learning algorithms typically produce a range of waveform type classes [1,2]. Here, WaveMAP [3], an unsupervised clustering framework that integrates graph-based clustering with nonlinear dimensionality reduction, was applied to extracellular recordings from the auditory cortex of awake common marmosets (Callithrix jacchus) using silicon probes. Single units were isolated using signal-to-noise ratio criteria, and spatiotemporal EAP kernels were extracted over a 180 um vertical span centered on the peak channel and a 1.5 ms temporal window surrounding each spike. A representative multi-channel waveform was generated for each single unit by averaging and normalizing its kernels, after which the vectorized spatiotemporal kernels were clustered using WaveMAP. The resulting clusters were classified as broad- or narrow-spiking based on their median waveform duration. WaveMAP identified six broad-spiking and four narrow-spiking waveform types that exhibit distinct waveform features and firing dynamics [Fig. 1]. These EAP types further differ in their laminar distributions, electrophysiological properties, and functional response characteristics, suggesting they represent distinct neuronal populations within the marmoset auditory cortex. Together, these findings demonstrate that unsupervised analysis of spatiotemporal waveforms captures physiologically meaningful diversity without existing neuron type labels. This framework provides a scalable approach for the comparison of cortical organization and the identification of conserved spatiotemporal patterns across species.
-
Authors: Reed Barmada, Rozita Laghaei, Stephen Meriney
Abstract: At the neuromuscular junction, a single action potential triggers a cascade that unfolds across many orders of magnitude in both space and time. Within milliseconds, the rising and falling voltage of the action potential gates presynaptic calcium channels, but whether a vesicle fuses depends on what happens next at the nanometer scale: how many calcium ions enter, where they diffuse, and whether they reach a calcium sensor protein binding site before diffusing away of being buffered. The timing of channel opening relative to the action potential waveform is not a detail that can accurately be derived by averaged kinetic events. The stochastic nature of this process shapes release probability, short-term plasticity, and the functional deficits seen in diseases like Lambert-Eaton Myasthenic Syndrome, where a subset of presynaptic calcium channels is lost to autoimmune attack. The computational tools that have historically handled these two scales, electrophysiology simulators such as NEURON for the action potential, and stochastic reaction-diffusion simulators such as MCell for calcium microdomains, were designed to work independently. Coupling them rigorously, without sacrificing temporal accuracy on one side or spatial resolution on the other, has remained an open problem. Many hybrid approaches sidestep this by driving MCell with simplified calcium pulses or fixed rate constants, which is fast but discards the very stochastic features of the process that matter most.
-
Authors: Danial Zenoozi, Trisha Basu, AmirEmad Ghassami, Edward Zagha, Erfan Nozari
Abstract: Effective connectivity identifies directional neural interactions, but widely used methods such as Granger causality rely on assumptions about linearity and stationarity, particularly when applied in the frequency domain. These assumptions are, importantly, scale dependent: cellular activity is strongly nonlinear, whereas spatiotemporal averaging can render macroscopic dynamics approximately linear. We develop a framework linking spatial scale, integrated state dynamics, and biological timescale to the reliability of time- and frequency-domain causal discovery using multivariate Granger causality. Our hypothesis is that time-domain causal discovery remains robust at microscopic scales, whereas frequency-domain causal discovery becomes reliable after sufficient averaging under stationary conditions. Furthermore, we anticipate that when spectral analysis is viable, the informative bands reflect the kinetics of the underlying interactions.
-
Authors: Xiaohan Zhang, Alex Doran, Zhe Sage Chen
Abstract: In recent years, digital twin brain (DTB) has emerged as an important research area that leverage the exponential growth of knowledge in brain connectomes, large-scale neural recordings, deep learning, and neurostimulation [1]. By modeling the whole brain or localized circuits in silico, the digital twin represents a biological replica, which may provide mechanistic insights into functions of healthy and diseased brains as well as personalized treatment such as deep brain stimulation (DBS). Motivated from the mouse behavioral and electrophysiological data [2,3], we have previously developed a biologically constrained, cell-type specific neural network as a digital twin for the mouse prefrontal cortex and mediodorsal (MD) thalamus in cognitive control and flexible decision making [4]. We showed that the addition of a feedforward MD structure (“controller”) to the recurrent PFC increases robustness to low cueing signal-to-noise ratio (SNR), enhances working memory, and enables rapid context switching. The model replicates key neurophysiological findings in task-performing mice, revealing computational mechanisms of MD in regulating cue uncertainty and context switching to enable cognitive flexibility [4]. We recently extended this model via adding two important features: (i) incorporating multiplicative gating mechanism [5]; (ii) generalizing rate-based to spiking-based neural networks. First, the spiking neural network model replicates the key findings of the prior rate-based neural network model. Next, we show that emergent rhythmic couplings emerge in a large-scale (N=~1000-2000 spiking neurons) task-performing PFC-MD network, where the LFP proxy may serve as a robust readout of interareal communication. Furthermore, different thalamic gating mechanisms (additive alone vs additive + multiplicative) yield qualitatively and quantitatively different thalamocortical LFP couplings in both amplitude and phase. Finally, from a control perspective, thalamic stimulation provides a means to modulate cortical dynamics in cognitive control and enhance behavioral flexibility of disease phenotypes. Our DTB model provides a platform for in silico MD stimulation for testing different open- or closed-loop thalamic stimulation strategies and providing experimentally testable hypotheses in animal models.
-
Authors: Balamurugan Ramachandran, Victoria Fouhy, Michael Smotherman, Rolf Muller
Abstract: Many bat species inhabit dense, cluttered environments where they must extract navigational information by analyzing returning clutter echoes. Prior work on biosonar stimulus ensembles has identified echo categories based on statistical invariants in these returns. Still, electrophysiological studies of bat biosonar have relied almost exclusively on simple, mostly deterministic stimuli such as pure tones and frequency-modulated sweeps. Whether the bat’s brain organizes its response to natural clutter echoes along the same acoustic dimensions that emerge from unsupervised statistical analysis remains an open question. We have investigated this question by a two-step experimental approach (Fig. 1) that first discovered echo categories from biomimetic sonar recordings in natural environments using unsupervised deep learning and then tests whether those categories can be decoded from neural activity. Biomimetic sonar echoes recorded along walking paths at a tropical forest site (Tasek Merimbundum, Brunei) were converted to spectrograms and encoded with a β-variational autoencoder [1] into a compact latent space. We clustered these latent codes with k-means to yield five distinct echo-type categories.
These five echo-type categories were presented as playback stimuli during cortical auditory evoked potential (CAEP) recordings from anesthetized bats. A residual network (ResNet) classifier [2], jointly tuned over trial-block averaging size and post-stimulus time cutoff, achieved 93.3% five-class decoding accuracy (Fig. 2c), demonstrating that the neural response carries reliable, category-specific information. To localize where in the CAEP trace this information resides, we identified a per-trial transition point from an early, variable segment to a later, stabilized segment using a rolling measure of signal variability, and compared decoding accuracy between the two windows; classification-relevant information is concentrated in the early window (Fig. 2a, 2b), a finding substantiated by Grad-CAM [3] attention maps that show strong activation before the transition and a sharp drop-off afterward (Fig. 3). These results demonstrate that the bat’s neural response to natural clutter echoes reflects the same dimensions captured by unsupervised acoustic clustering. Ongoing work will extend the stimulus set to estimated impulse responses, which are echoes convolved with the bat’s own emitted chirp, to test whether a more naturalistic stimulus presentation elicits a stronger or more differentiated cortical response.
-
Authors: Liang Chen, Shir R. Maimon, Nachum Ulanovsky, Johnatan Aljadeff
Abstract: We aim to understand population-level spatial representations of naturalistic environments and how spatial coding interacts with locomotion and trajectory planning in hippocampal area CA3. We extend our previous mean-field approach [1] to derive a spatial attractor model from a recurrent network of Izhikevich spiking neurons [2] (Fig. 1A). Unlike classical rate-based attractors, this framework links heterogeneous neuronal properties, nonlinear subthreshold dynamics, and spike-frequency adaptation (SFA) to population-level spatial representations. Neuronal heterogeneity generates a broad distribution of place-field sizes, reproducing the multiscale spatial coding observed experimentally in large environments [3, 4] (Fig. 1B) Animal movement further reshapes these representations: place-field sizes and centers are significantly modulated by movement speed (Fig. 1C). We identify a network ‘speed limit’ that determines whether the population activity bump can reliably track the animal’s position (Fig. 1D). Below this limit, the bump remains locked to the moving sensory input; above it, tracking fails. The speed limit is derived analytically from parameters controlling bump size and increases with neuronal heterogeneity, predicting how cellular diversity supports robust spatial coding during rapid movement. Consistent with analysis of experimental data from [4], we propose two movement-dependent coding regimes (Fig. 1E). At lower speeds, neurons experience relatively homogeneous excitability, producing a narrower distribution whose characteristic scale increases with movement speed. At higher speeds, increased heterogeneity supports reliable tracking while generating a substantially broader field-size distribution with weak speed dependence. Thus, movement and heterogeneity jointly determine both individual place-field scales and the population coding regime. The model reproduces two experimentally observed transient modes for updating an internal position estimate after sudden relocation of sensory input [5] (Fig. 1F). Depending primarily on recurrent spatial coupling, the activity bump either rapidly relocates to the new position or continuously propagates through intermediate positions, effectively “simulating” movement toward the new location. This propagating regime links attractor dynamics to internally generated transient neural sequences. Finally, SFA provides an intrinsic mechanism for sequential activity (Fig. 1G). Without external input, sufficiently strong and directionally biased adaptation destabilizes a stationary bump and drives spontaneous propagation, reproducing replay-like sequences during immobility or sleep. During locomotion, competition between intrinsic adaptation and moving sensory input generates biased forward and backward sequential activity, resembling hippocampal theta sweeps [6]. Together, the model links cellular heterogeneity and adaptation to multiscale place coding, movement dependent spatial representations, tracking limits, and sequential activity, providing analytically testable predictions for large-scale hippocampal recordings
-
Authors: Peter Dominey, Jean-Marc Fellous, Alfredo Weitzenfeld
Abstract: When navigating in complex environments, decisions are made on the basis of past information. That information can be temporal (e.g. 10 seconds ago) or spatial (e.g. back 2 meters). Our work discusses the influence of past spatial and/or temporal information on the coding of prefrontal cortex (PFC) neurons in the rodent during spatial navigation in large (“megaspace”) environments by analyzing predictions from a reservoir computing network, a type of sparsely connected recurrent neural network that has been previously used to model the primate PFC. Neurons in recurrent reservoir network models and in experiments have been shown to display high dimensional mixtures of spatio-temporal task relevant parameters. The goal of this research is to understand and model the learning of spatio-temporal representations in rodents and their ability to integrate serial and temporal structure during spatial navigation in which speed is manipulated to tease apart the contribution of temporal and spatial past information. Our hypothesis is that a reservoir network can predict and explain mixed spatial and temporal selectivity during learning. This investigation is novel and unique in trying to understand how the interactions between spatial and temporal processes facilitate navigation learning and decision making in large spaces. This research involves the testing of predictions from a reservoir computing model using experimental studies that use a novel rat-robot co-navigation paradigm. In this unique paradigm, the temporal structure of behavior can be precisely controlled by manipulating speed in a large open environment during natural navigation. Our initial findings show that rodents are able to use speed as a cue for spatial decision making. The proposed reservoir model provides a new tool for the detailed study of the interaction of time and space in PFC during spatial decision making in large environments. In this context, the powerful concept of mixed selectivity is extended into the temporal domain. This research will contribute to computational neuroscience along two different axes: (1) using data-constrained reservoir computing to propose mechanisms explaining the nature of temporal and spatial interactions during complex navigation and (2) integrating the reservoir sequence-producing PFC model with and existing hippocampus spatial representation model to provide a system-level view of complex spatial navigation appropriate for robotic NeuroAI implementations.
-
Authors: Runnan Cao, Jie Zhang, Yilin Li, Yue Wang, Hohyun Cho, Peter Brunner, John R. Pruett, Jr., Jon T Willie, Jasmine Thum, Shuo Wang
Abstract: First impressions form within a few hundreds of milliseconds of seeing a face, requiring a rapid transfor- mation from visual representation of facial features to abstract social evaluations 1. How this sensory-to-social transformation unfolds across the hierarchy, and how it is disrupted in conditions of atypical social cognition, remain poorly understood. These judgments are thought to arise from the visual analysis of facial features 2, yet how the human brain transforms a visual representation of a face into abstract social evaluations has remained unresolved. Here we recorded intracranial EEG across a large-scale face-processing and social-evaluative network in 31 neurosurgical patients, together with 1,909 single neurons recorded in 30 patients, while participants viewed 500 natural face images. Rating across 8 different social traits, which covers a comprehensive four-dimensional trait space were obtained from a large cohort of participants in our previous study.
-
Authors: Victoria Fouhy, Sam Ellis, Michael Smotherman
Abstract: Echolocating bats and cetaceans are ideal for investigating mammalian auditory processing and sensorimotor integration because they use temporally structured vocal sequences as a source for continuous, behaviorally relevant sensory feedback. Unlike most animal vocalizations, which primarily communicate identity, fitness or intent, echolocation signals encode rich contextual information about the surrounding environment. Experimentally manipulating pulse-echo delays enables precise control over the bat’s perceived acoustic scene, allowing for systematic alterations of sensory input in an ethologically meaningful way while maintaining an unambiguous relationship between stimulus and perception. Here, we aim to identify where and how in the brain neurons selectively respond to ecologically significant patterns of change derived from natural acoustic sequences.
-
Authors: Giacomo Vedovati; Thomas Papouin; ShiNung Ching
Abstract: Astrocytes are non-neuronal cells of growing computational interest due to their capacity to gate and bias neuronal circuit activity and dynamics over extended timescales [1]. In this work, we construct and study a hierarchical fast-slow neural-astrocyte architecture where a slow astrocytic network modulates a fast neuronal one during a multi-context, feedback-driven bandit-like task. We show that this network implements a win-stay/lose-switch hybrid automaton, where discrete contextual transitions are enacted via astrocyte-mediated continuous evidence accumulation. Dynamically, this evidence integration is governed by bifurcations that deform the astrocytic vector field in response to reward distribution/omission, inducing low-velocity regions whose traversal corresponds to the agents’ stickiness.
-
Authors: Danna Pinto, Dan Harel, Robert Knight, Peter Brunner, Leon Deouell
Abstract: Neural representations of visual object categories are well established throughout the visual processing hierarchy, with robust category-specific representations in higher-order ventral temporal cortex [1]. Recent studies have further shown that category-specific representations in higher-order visual cortex can remain stable throughout prolonged stimulus presentation, extending well beyond transient responses to stimulus onset [2,3,4]. Do these sustained representational dynamics persist when attention is directed away from the visual object?
-
Authors: J. J. Wallner, M. Angrick, S. Luo, D. Candrea, N. E. Crone, D. J. Krusienski
Abstract: Stable long-term decoding remains a critical barrier to the clinical deployment of intracranial brain– computer interfaces (BCIs), including electrocorticographic (ECoG) systems for people with paralysis and chronic use in ALS [1,2,3,4]. High-gamma activity (HG, >70 Hz) dominates speech and motor decoders but is susceptible to slow signal drift that degrades performance over months to years and forces frequent recalibration [5,6,7,8]. Lower-frequency oscillations also carry movement- and speech-related information and are more spatially distributed than HG, raising the question of whether multi-band features can improve longitudinal decoding stability without sacrificing accuracy.
-
Author: Yan Yan
Abstract: Volumetric fluorescence microscopy characterizes neural tissue across complementary spatial scales. Gene-expression signals appear as numerous subcellular puncta; cells and nuclei form compact, frequently touching structures; and neuronal axons and dendrites extend as sparse or densely intertwined arbors over long distances [1,5]. Recovering individual biological entities at all three scales would support assignment of molecular signals to specific cells and association of cellular identities with larger-scale neuronal morphology. The structures nevertheless occupy markedly different geometric regimes, and most analysis pipelines are optimized for only one morphology or output representation.
-
Authors: Yahya Shirazi; Aviv Dotan; Bruno Aristimunha Pinto; Oren Shriki; Amitava Majumdar; Scott Makeig; Arnaud Delorme
Abstract: EEG-DaSh (EEG Data Sharing) is a CRCNS project in its second year that is building a data and tool sharing resource designed to accelerate discovery in human neuroelectromagnetic research. A United States–Israel collaboration between the Swartz Center for Computational Neuroscience at the University of California San Diego and Ben-Gurion University of the Negev, EEG-DaSh builds on the success of EEGLAB, the NIH BRAIN-funded OpenNeuro and NEMAR (Neuroelectromagnetic Data Archive and Tools Resource) archives, and the Neuroscience Gateway to provide the first online platform that pairs curated, machine learning (ML) and deep learning (DL) ready electroencephalography (EEG), magnetoencephalography (MEG), intracranial EEG (iEEG), and electromyography (EMG) datasets with open pipelines for analysis, modeling, and data augmentation at archive scale. EEG-DaSh delivers this through two coordinated components, the NEMAR archive and the EEGDash software package, described in turn below.
-
Authors: Samuel A. Nastase, Dani Byrd, Shrikanth Narayanan
Abstract: We spend much of our lives in ongoing dialogues with the people around us. How does the brain drive these dialogues forward? Why do we say what we say? The spontaneous, intentional, and contextual nature of everyday conversations presents a fundamental challenge to the experimental paradigms of neuroscience and psychology. Most research on the neurobiology of language has focused on more the more tractable problem of language comprehension, or on simple, highly constrained production tasks like picture naming or sentence repetition. In real-world dialogues, our words at any given moment are driven by the current conversational context and the shared history of conversations with a particular person—a history that could span minutes, days, months, or even years. How can we begin to model this phenomenon?
-
Authors: C. Giossi; J. Rubin; T. Verstynen; C. Vich
Abstract: Adaptive behavioral control requires continuous calibration of action speed and the capacity to suppress responses when facing unexpected stop signals. In computational models, these two forms of adaptation are captured by distinct parameters: adjustments in the evidence accumulation rate (drift rate) and modifications in the decision threshold (boundary height). While phenomenological models successfully fit these parameters to human data, how these distinct adjustments are physically instantiated within biophysically constrained basal ganglia networks remains a major open question. Specifically, it is unclear whether a uniform, dopamine-dependent corticostriatal plasticity rule is sufficient to jointly regulate both movement vigor and adaptive threshold changes, or whether they rely on dissociable circuit mechanisms.
-
Authors: Daisuke Yamada; Qijun Zhang; Travis Pence; Barbara B. Bendlin; Federico Rey; Vikas Singh
Abstract: Compositional data—vectors encoding relative proportions—arise across scientific domains, including ecology, geochemistry, and genomics. The features in these data often come with known hierarchical structure (e.g., taxonomies, phylogenies, ontologies), yet existing methods either ignore this structure, discard the intrinsic Aitchison geometry, are designed for binary trees, or yield incomplete coordinate systems. We describe Poly-ILR, a canonical orthonormal decomposition of the Aitchison tangent space aligned with any tree topology. Our construction defines a weighted local geometry at each internal node capturing full branching structure, then lifts these to a global orthonormal basis where every coordinate corresponds to a specific tree location. On microbiome and single-cell benchmarks, PolyILR yields stable, interpretable features and enables inference at multiscale tree resolution. We also establish a novel theoretical connection to softmax classifiers, suggesting possible applications to probabilistic modeling.
-
Authors: Johannes Bertram; Luciano Dyballa; T. Anderson Keller; Savik Kinger; Steven W. Zucker
Abstract: A central goal in systems neuroscience is to compare computations across neural systems — brain regions, organisms, or deep learning models. The dominant approach probes each system with a shared set of stimuli and records the population response — the state — for each stimulus. These states are then compared using scalar alignment scores such as Representational Similarity Analysis (RSA) and Centered Kernel Alignment (CKA). While states are useful for decoding — the stimulus can be inferred from the population activity — scores over states offer only a partial view of computational organization. How many neurons actually drive the observed state? How unique is it, and how dependent on the training objective? We show: Two systems that are not “digital twins” can achieve indistinguishable alignment scores — yet the same population state can be produced by qualitatively different internal functional organizations.
-
Authors: Kristin Schmidt; Fabrizio Sergi
Abstract: Humans can skillfully interact in uncertain and dynamic environments, yet how the nervous system achieves this robustness remains poorly understood. One mechanism that supports these interactions is the regulation of joint impedance through the simultaneous activation of muscle groups, or muscle coactivation. However, it remains unclear whether coactivation emerges from the same neural processes that generate force or recruits additional neural resources dedicated to impedance regulation. Resolving this question is critical for advancing models used in motor control, rehabilitation, and robotics.
-
Authors: Isabel Fernandez, Maxwell Greene, Vinith Kugathasan, Ramón Bartolo, Jan Antolík, Bevil. R. Conway, Daniel A. Butts
Abstract: Visual inputs are inherently temporal – shaped by self-motion and eye movements – and the timing of neural responses in the primary visual cortex (V1) reflects both spatiotemporal selectivity and their position in the visual circuit. Yet data-driven deep neural network (DNN) models of V1 typically neglect temporal processing, either by training based on flashed-image responses or by confining spatiotemporal selectivity to the input stage. Here, we demonstrate that fitting data at high temporal resolution and distributing temporal selectivity across many stages of a DNN can provide useful insights into V1 processing. We use a DNN-based approach called computational barcoding to characterize the computational role of neurons within V1 populations and apply it to both data generated from a large-scale realistic dynamical simulation – with known per-neuron computations – and real V1 data. We show that distributing temporal processing across the model results in barcodes that better distinguish functional differences across the V1 population. Likewise, representing and fitting the data at higher temporal resolution improves performance. In the simulated data, we confirm that these modifications more accurately recover known computational distinctions. This suggests that high-resolution temporal models are useful in understanding visual processing, which is often considered only spatially.
-
Authors: Osvaldo M Velarde; Lucas Parra
Abstract: Human vision is fundamentally active. The brain samples the environment through rapid eye movements to shift eye fixation. How these snapshots are integrated over time remains an open question. We propose the hypothesis that biological vision uses recurrent feedback to integrate information across fixations and is tuned to predict the content of the next fixation. In the literature, computational models of saccadic vision have mainly focused on predicting where humans look. Here, we address what can be inferred from a sequence of fixations by learning visual representations from saccadic sequences in a self-supervised approach, namely prediction of the content of the next fixation with a recurrent integration of the previous fixation representations. Then, to measure how well the gist of a scene is captured with this recurrent system, we translate the integrated visual context into human text captions of images, using techniques and data from the image-captioning literature.
-
Authors: David St-Amand; Greg D. Field; John Pearson
Abstract: The efficient coding hypothesis holds that early sensory systems should compress their inputs before passing them to the rest of the brain. In the retina, efficient coding predicts that because most of the power in natural movies sits at low spatial and low temporal frequencies, RGC types that jointly tile spatio-temporal frequency space must trade spatial resolution against temporal resolution: small receptive fields pair with extended temporal integration and vice versa. This is broadly consistent with observed RGC diversity, with one exception: blue-yellow color-opponent RGCs have large receptive fields and slow temporal integration, low spatial and low temporal frequency preference at once. Prior work has treated chromatic and spatio-temporal coding separately; no study has asked what happens to the predicted RGC organization when color, space, and time are optimized jointly. We address this gap directly.
-
Authors: Noam Joshua, Adva Danielly, Shany Margalit and Arseny Finkelstein
Abstract: Everyday behavior depends on remembering where important places and resources are located, even when they are out of sight and unrelated to ongoing actions. Yet it remains unclear whether the brain maintains an internal representation of where value is located, how it is distinguished from current location representation and planned movement, and how it is updated when contingencies change. We developed a large, multi-target maze with eight targets organized into four spatial patches. During each block, one patch contained the two most valuable targets, its identity changed across the session, requiring animals to discover, remember, and update the valuable location. We used Chronic Neuropixels recordings to target frontal cortical areas. Individual neurons exhibited transient and sustained selectivity for target location, patch, and reward outcome. Population activity predicted the animal's current target and patch several seconds before arrival, consistent with prospective planning.
-
Authors: Adam Hinson; Rolf Müller
Abstract: Bat biosonar operates on a sensory stimulus ensemble fundamentally different from that of vision. In vision, objects and surfaces generally have a comparatively direct and stable spatial correspondence to locations within the visual field. In biosonar, however, each brief emission interacts with numerous leaves, branches, and other surfaces, producing a one-dimensional echo waveform composed of overlapping reflections. Small changes in position, orientation, or the arrangement of surrounding clutter can reorganize these reflections and produce large, difficult-to-predict changes in the received signal, even when the broader environment remains unchanged. Despite this highly variable sensory input, bats can fly rapidly through dense vegetation, avoid obstacles, detect and classify objects, pursue moving prey, and maintain spatial orientation [1]. Their ability to extract reliable, behaviorally useful information from such an unstable acoustic stimulus suggests that biosonar perception depends on representations capable of revealing structure hidden within the temporal dynamics of echoes.
-
Author: Caterina Stamoulis
Abstract: Task-independent (intrinsic) brain oscillations play a fundamental role in cognitive function. However, it is unclear whether infra-slow (<0.5 Hz) BOLD oscillations encoded in resting-state fMRI signals from the developing brain can be robustly estimated. In addition, their spatiospectral characteristics, structural underpinnings and cognitive correlates remain incompletely understood. In N=3507 adolescents (median(IQR) age =12.0(1.1) years; 50.9%-females) from the Adolescent Brain Cognitive Study, intrinsic oscillations were estimated from resting-state fMRI with an unsupervised time-domain method.
-
Authors: Zachary M. Miksis; Peter Jedlicka; Alexander Opitz; Andreas Vlachos; Gillian Queisser
Abstract: Neurons rely predominantly on electrical and biochemical signaling to encode information and adapt for development, learning, and survival. Neuronal calcium signals play an outsized role in regulating information transfer from synaptic sites to the cell nucleus. On the one hand calcium signaling regulates biochemical cascades relevant for learning, plasticity, and cell survival. On the other hand, disruption of calcium signals is implicated in neuronal decay and onset of clinical symptoms. In the context of transcranial magnetic stimulation (TMS), a non-invasive brain stimulation method used in the treatment of certain neuropathologies, such as depression and schizophrenia, the modulatory effects of TMS on intracellular calcium dynamics is not well understood. In a collaborative project around the development of the multiscale simulation toolbox “Neuron Modeling for TMS” (NeMo-TMS) we have been integrating highly detailed calcium models, including intracellular calcium store dynamics, to study the effects of different repetitive TMS (rTMS) brain stimulation protocols on intracellular calcium signaling and the potential long-term adaptive properties of rTMS through calcium signals. Our computational results demonstrate the relevance of store-activated calcium during TMS treatment. We further show that pause length between TMS pulses is a parameter that can be used to maximize cellular calcium responses and that the timing and synaptic location of pre-synaptic and post-synaptic TMS-induced activation controls the intracellular calcium response. Taken together these findings provide guidance towards optimizing rTMS applications.
-
Authors: Harman Dewantoro, Catherine Stamoulis, Prodromos Daoutidis
Abstract: Intrinsic brain connectivity is typically estimated using correlation measures that assess pairwise associations between regional blood-oxygen-level-dependent (BOLD) signals. Resulting connectivity graphs are used to assess the brain’s functional organization [1]. A shortcoming of estimated correlations is that they represent both direct associations of intrinsic activity between pairs of regions but also indirect relationships resulting from connections to other common regions[2]. To decouple these direct and indirect relationships, previous causal-discovery studies have examined resting-state brain connectivity using score-based [3], constraint-based [4,5], non-Gaussian structural-equation [6], and hybrid approaches [7]. Some approaches assume acyclic networks [3,5], while others rely on non-Gaussian signal structure to assign edge directions [6,7], without modeling temporal dependencies.
-
Authors: Reed Barmada; Rozita Laghaei; Stephen Meriney
Abstract: At the neuromuscular junction, a single action potential triggers a cascade that unfolds across many orders of magnitude in both space and time. Within milliseconds, the rising and falling voltage of the action potential gates presynaptic calcium channels, but whether a vesicle fuses depends on what happens next at the nanometer scale: how many calcium ions enter, where they diffuse, and whether they reach a calcium sensor protein binding site before diffusing away of being buffered. The timing of channel opening relative to the action potential waveform is not a detail that can accurately be derived by averaged kinetic events. The stochastic nature of this process shapes release probability, short-term plasticity, and the functional deficits seen in diseases like Lambert-Eaton Myasthenic Syndrome, where a subset of presynaptic calcium channels is lost to autoimmune attack. The computational tools that have historically handled these two scales, electrophysiology simulators such as NEURON for the action potential, and stochastic reaction-diffusion simulators such as MCell for calcium microdomains, were designed to work independently. Coupling them rigorously, without sacrificing temporal accuracy on one side or spatial resolution on the other, has remained an open problem. Many hybrid approaches sidestep this by driving MCell with simplified calcium pulses or fixed rate constants, which is fast but discards the very stochastic features of the process that matter most.
-
Authors: Tadashi Yamazaki, Kaaya Akira, Rin Kuriyama, Towa Muranaka, Hinari Sugimoto
Abstract: Neulite is a lightweight neuron simulator designed for biophysically detailed neuron and network models [1]. It works closely with the Brain Modeling ToolKit (BMTK), a Python package developed by the Allen Institute [2]. Models built with BMTK's "bionet" module and executed on the NEURON simulator can run on the Neulite kernel by using the "bionetlite" module, which serves as a drop-in replacement for bionet. Therefore, the NEURON-based and Neulite-based implementations within BMTK should yield identical simulation results. In this study, we examined the reproducibility of simulation results between these two versions. However, while Neulite currently supports only neuron models registered in the Allen Cell-Types Database, NEURON accommodates a much wider variety of models, such as the mouse primary motor cortex model [3]. To address this limitation, we also explored methods for implementing non-BMTK models on the Neulite kernel, aiming to ensure interoperability between NEURON and Neulite. To evaluate reproducibility, we used the Allen V1 model [4] and compared the firing rates of individual neuron populations in response to constant current inputs. By aligning the temporal resolution and numerical methods, we obtained quantitatively comparable results. Furthermore, we found that performing parameter fitting enhanced the reproducibility of the voltage traces. Regarding interoperability, we leveraged large language models (LLMs) to translate NMODL files into C code for Neulite and to perform parameter fitting, which yielded qualitatively comparable results for individual neurons in the mouse primary motor cortex. In conclusion, although additional parameter tuning is required, Neulite is capable of generating numerical results close to those of NEURON.
-
Authors: Masaharu Yonezawa, Tadashi Yamazaki
Abstract: Repetitive transcranial magnetic stimulation (rTMS) is widely used as a non-invasive neuromodulation approach for neurological and psychiatric disorders, but how local cortical circuit activity changes over time under different stimulation frequencies remains incompletely understood. Most modeling studies have focused on single-pulse stimulation or single-neuron responses, leaving circuit-level analyses that integrate morphological detail, synaptic connectivity, and repetitive stimulation limited. Here, we incorporated TMS- induced electric fields into a morphologically detailed local circuit model of mouse primary visual cortex (V1) to examine how rTMS frequency shapes layer- and cell-type-specific activity over time.
-
Authors: Satwik Pasani; Xiaoxuan Lei; Dean Pospisil; Benjamin Cowley
Abstract: Computational models that predict neural activity given visual stimuli provide a powerful modality to understand computations underlying visual cortical activity. The current state-of-the-art performance is achieved by task-driven Deep Neural Networks (DNNs) trained to perform tasks such as image classification. This requires regressing high-dimensional (≈ 50 – 100k) feature vectors from intermediate layers of these DNNs against limited neural data. Current performance estimates have a substantial predictive gap; for example, task-driven DNN features achieve R² ≈ 0.3 – 0.5 for V4 neuronal responses to natural images (Fig 1A). How do we bridge this predictive gap? There are three broad possibilities. The DNN features could be missing computations important for neural activity prediction – we would need more biologically aligned models that capture the relevant computations. Alternatively, we could be using too few DNN features – by increasing the number of features (potentially concatenating across layers or different DNNs) we will be able to surpass this ceiling. Lastly, regularized regression on noisy neural responses leads to biased performance estimates – more training data for the linear model would improve performance. In this work, we disambiguate between these possibilities by introducing a novel metric, R̂²=, that estimates the proportion of variance explained by an optimal linear model (i.e. a model trained with infinite training data) over DNN features even when the number of features is much larger than the number of training samples.
-
Authors: Michael Finch, Michael Vivian, Lisandro Martin, Christine Grienberger, Aaron D. Milstein
Abstract: During spatial learning, hippocampal place cells rapidly change their selectivity for features of the environment, with a bias towards over-representing goals. These adaptations are driven by a recently discovered form of plasticity induced by dendritic calcium spikes termed behavioral timescale synaptic plasticity (BTSP)1. Dendritic calcium spikes are promoted by dendritic excitation by long-range “teaching signals” from cortex, and limited by local dendritic inhibition2,3. Here we investigate the relative contributions of feedback SST+/Chrna2+ (OLM) and feedforward NDNF+ dendrite-targeting interneurons to place field formation and goal over-representation in CA1 by analyzing in vivo calcium imaging data (Fig. 1A). Latent factor analysis4 reveals that OLM and NDNF+ interneurons functionally segregate into multiple subpopulations with distinct profiles of modulation by reward and adaptation over learning (Fig. 1B–I and Fig. 2). Circuit modeling suggests complementary roles for OLM and NDNF+ cells – suppression of OLM interneurons at rewards disinhibits pyramidal cell dendrites and promotes plasticity at reward locations, while enhancement of NDNF+ interneuron activity at rewards provides a brake on plasticity that may limit the number of place cells recruited to represent a rewarded location (Fig. 3). These results support important roles for feedforward, feedback, and disinhibitory components of the dendritic inhibitory microcircuit in regulating spatial memory formation.
-
Authors: Mark A. Reimers; Michael G. Moore
Abstract: We have developed an inexpensive approach to wide field calcium fluorescence imaging at high (single-cell) spatial and temporal (50ms) resolution. Using this system we have recorded activity from up to 1,000 neurons simultaneously over the entire dorsal cortex of a mouse. We have verified that the identified units in visual cortex show tuning characteristic of single cells, unlike related efforts [Kauvar]. The apparatus is put together from inexpensive off-the-shelf components. What makes it work is an adaptation of recent developments in optimization theory, in particular the Alternating Direction Method of Multipliers (ADMM). We apply this ADMM method to solve a heavily constrained and parallelized version of the Non-negative Matrix Factorization algorithm.
-
Author: DJ Hamilton
Abstract: Toward understanding the molecular mechanisms of human cognition, the transformer architecture has provided useful insights via analogous behaviors observed in some Large Language Model (LLM) instantiations. And, because quantum physics constrains cognitive behaviors, the associated mathematical apparatus of quantum mechanics, including constructs such as superposition and entanglement, can be used to rigorously model cognitive processes. The advent of quantum transformers combines both transformer and quantum aspects into an enhanced LLM capability. Quantum transformer-based LLMs (QLLMs) provide an even closer analogy to human cortical processing than is currently provided by classical implementations. Leveraging QLLM's accessibility relative to healthy human brains facilitates our ability to significantly improve our understanding of cognitive processing.
-
Authors: Shizhao Liu, Anton Pletenev, Ralf M. Haefner, Adam C. Snyder
Abstract: Computational question. Does learning improve sensory representations by reducing redundancy, or by sharing information among neurons? Models that optimize sensory representations for downstream discrimination often predict reduced information-limiting correlations. Hierarchical generative inference instead predicts that learning the statistical structure of a task increases shared information: sensory neurons represent posterior beliefs that combine local evidence with information propagated through the network. We tested these competing predictions using population recordings in macaque visual cortex and simulations of hierarchical Bayesian inference.
-
Author: Paola Malerba
Abstract: Sleep is a state of highly rhythmic brain dynamics contributing to health and cognition. Studies in humans typically examine sleep brain oscillations measured by electroencephalography (EEG), and have shown that sleep oscillations serve to organize information processing and communication in relation to a specific behavior, and are related to health functions of sleep such as homeostatic maintenance of synaptic architecture and clearance of the glymphatic system. Yet, the fundamental properties of brain oscillations that mediate these relations to function are not understood. This underscores the need for generalizable analytical methods that can support development of mechanistic knowledge linking sleep EEG to functional cognitive and health outcomes. This needs to be paired with understanding of within-subject (ie., personalized) factors influencing how changes in space-time organization of the sleep EEG can act as markers of functional differentiation, for example in translational studies.
-
Authors: Christopher S. Peters, Kristin Schmidt, Fabrizio Sergi
Abstract: In tasks involving physical interaction, humans need to account for uncertainty about the dynamics of the en- vironment or the interacting agent. While energetically costly, co-contraction of antagonistic muscle pairs is used to increase joint impedance during physically interactive tasks, leading to successful task performance even when the environment dynamics are unstable or unpredictable [1]. Stochastic optimal control (SOC) algorithms allow to predict the optimal motor commands for a physically interactive task by trading-off task performance with control effort in the presence of uncertainty [1]. Paired with an accurate muscle model, SOC can be used to study the optimal control of muscle force and impedance in simulation.
-
Authors: Leonardo E. S. Tavares, Benjamin M. Zemel, Madza Farias-Virgens, Rodrigo F. O. Pena, Claudio V. Mello, Henrique von Gersdorff
Abstract: Avian brains control fast, precise, and complex behaviors using 3-fold less energy per neuron than mammalian brains. How is this high performance achieved with such low energy cost? One intriguing mechanism links higher avian brain temperature (~40°C) to energy efficiency. To address this question, we developed a biophysical model that recapitulates in vitro recordings from song motor control neurons in zebra finches (Figure 1A). This model exhibits an exquisite coordination of ionic currents at 40°C that minimizes Na+ influx during spike firing (Figure 1B). We then used in vivo spike recordings from singing birds to model the temperature dependence of their burst-pause firing statistics (Figure 1C). We found that elevated temperatures facilitated higher firing frequencies and discovered a remarkable tradeoff that rendered song energy consumption only weakly dependent on temperature from 30°C to 40°C (Figure 1D). Importantly, the >3-fold additional energy cost during song production compared to silent periods stemmed mainly from the plastic postsynaptic currents required to drive burst spiking and, surprisingly, not from spike firing per se (Figure 1E). The high physiological avian brain temperature thus enables dense, precise and energy-efficient spike codes in the specialized cortical neurons that drive song production.
-
Authors: Selina Li, Nathaniel Kinsky, Bapun Giri, Asohan Amarasingham , Kamran Diba
Abstract: Sleep is essential to memory consolidation, but the precise patterns of synaptic modification in this process remain unknown. Here, we analyzed a large-scale long-duration dataset from the hippocampus CA1 neuron in seven rats during sleep and sleep deprivation following novel maze exploration. We utilized cross-correlogram analysis to identify and categorize thousands of in vivo synaptic connections between pyramidal cells and interneurons and examine changes in their efficacy over time. Preliminary results focused on excitatory pyramidal-interneuron pairs revealed that while individual connections fluctuated, the overall strength of these connections was largely stable during undisturbed sleep. In contrast, we observed a significant increase in connection strength during sleep deprivation. Remarkably, the strength of these connections dropped substantially upon the onset of recovery sleep and continued to decrease further with additional sleep. These results indicate that while pyramidal-interneuron synapses remain stable during natural sleep, with both individual increases and decreases, they largely increase during forced wakefulness but with a strong decrease in the subsequent recovery sleep. Overall, these results appear largely consistent with predictions of the synaptic homeostasis hypothesis and highlight important distinctions between undisturbed sleep and recovery sleep following sleep loss.
-
Author: Chao Chen
Abstract: The cerebrovascular network is essential to brain function and is increasingly implicated in neurodegenerative and cerebrovascular disease. Light-sheet fluorescence microscopy (LSFM) now enables whole-brain imaging of the mouse vasculature at high resolution, producing volumetric datasets of unprecedented scale. Yet quantitative analysis of vascular structure remains largely limited to geometric descriptors such as vessel diameter, length, and branching density, which capture local morphology but are insensitive to the global organization of the network. Topological features such as connectivity, loop structure, and the hierarchical nesting of vessel trees encode biologically meaningful information that geometry alone cannot recover, but robust, scalable tools for extracting and comparing these features across brain regions and across animal populations are currently lacking. We present TopoVess, a topology-guided vessel analysis toolkit enabling principled, population-level comparison of vascular architecture in LSFM data.
-
Author: Ralf Haefner
Abstract: How sensory neuron activity relates to perception remains a major open question. Choice-related signals, the correlations between sensory responses and perceptual reports, are the field’s main window into this relationship, yet their magnitude varies widely across studies and neurons for unresolved reasons. Theory relates these signals to task-aligned correlated variability, which was recently shown to increase robustly in V4 as monkeys learned a task. Using the same recordings (two macaques, two orientation discrimination tasks, weeks of learning), we separated population choice signals into a task-aligned component, which existing models predict should grow with learning, and a task-orthogonal component of unknown origin.
Consistent with probabilistic inference models, the learning-related increase in covariability arose mainly within rather than between choices. This implicates graded feedback, such as task-related beliefs, rather than a binary choice signal. The increase also tracked task-aligned choice signals in three of four conditions. And as predicted, task-aligned choice signals grew with learning in the oblique task in both monkeys.
-
Authors: Sophia Nietubicz, Chenyang Lu, Julius Benson, Sarah Kim, Zhiyu Kuang, Daryl Meling, Howard Gritton, Jozien Goense, Dave Zhao, Paul J. Bonthuis
Abstract: Chronic social defeat stress (CSDS) is a widely used mouse model of stress-induced depression-like behavior, in which mice are typically classified as resilient or susceptible based on a strict cutoff for a single social interaction metric following defeat stress called the social interaction ratio (SI). However, this binary definition overlooks pre-stress baseline behavioral differences, reduces resilience to a single dimension, and obscures continuous variation in the degree and nature of social interaction. To address these limitations, we developed a multidimensional measure of the effects of social defeat stress by applying principal components analysis (PCA) to a panel of behavioral measures collected before and after CSD. We term this approach the longitudinal social interaction test (LSIT). Our first component captured a generalized avoidance tendency, characterized by increased time in corner zones located far from a social stimulus and reduced time in the social interaction zone near a stimulus, both before and after CSD. Our second component reflected stress-induced behavioral change, captured as the difference in sociability before versus after CSD. Both dimensions more effectively distinguished CSD-exposed mice from controls than the conventional post-CSD SI ratio. We then examined associations between these dimensions and resting-state fMRI connectivity following CSD. Whereas the standard post-CSD SI ratio measure showed no significant relationship to connectivity among brain regions implicated in social defeat stress, both PCA-derived dimensions were associated with multiple region-pair connectivity relationships (FDR < 0.10. We are currently developing BrainTree, a method that uses brain subnetworks to predict a mouse’s behavioral outcomes. BrainTree builds extremely randomized trees by sampling a subnetwork at each split via random walk with transition probabilities determined by edge Z-scores. On preliminary human fMRI data, it outperforms connectome-based predictive modeling (CPM) in predicting fluid intelligence, which helps us to identify possible predictive subnetworks in our mouse data. These findings suggest that resilience is best conceptualized as a continuous, multidimensional process shaped by individual behavioral trajectories rather than a static post-CSD outcome. By quantifying latent dimensions of stress trajectories, this new framework may improve identification of neural circuits underlying vulnerability and resilience to stress-related mental illness.
-
Authors: Behnaz Khaghani; Patrycja Oleniac; Astrid Lawyer; Hengji Huang; Jongwoon Kim; Yue Liu; Reuben Weitzman; Aliyah Oakley; Saphie Abdi; Shan Jiang; Xiaoting Jia; Harald Sontheimer; Meimei Liu
Abstract: Seizures engage distributed brain networks, but the region with the clearest local seizure signature need not be the one that changes first, and strong local detectability says nothing about how directed interactions between regions change at onset [1]. We asked whether regional seizure-state information, transition timing and directed phase-based coupling tell the same story when measured simultaneously.
One probe, four regions. A thermally drawn multifunctional fiber carries four BiSn electrodes, optical waveguides and a microfluidic channel [2]. Electrode windows at four depths sample cortex (CRX), hippocampal CA1 and CA3, and amygdala (AMG) along one insertion (tip ∼150 μm; ∼0.25 MΩ at 1 kHz). Only the electrodes were used here. We analysed LFPs from 13 awake, head-fixed mice before and during acute pilocarpine-induced seizures, using the first seizure per session and 5 s windows.