Periodic Reporting for period 2 - neuronsXnets (Network Analysis in Neocortex during Passive and Active Learning)
Berichtszeitraum: 2023-09-01 bis 2025-08-31
How does the brain compute to enable learning and interaction with the environment? Rapid advances in optical imaging, statistical/machine-learning methods, and computational resources create a timely opportunity to address this question. NeuronsXnets builds an international, multidisciplinary, intersectoral network spanning neuroscience, neuromorphic computing, data science, systems, optical imaging, and hospitals. It leverages this environment to deepen understanding of neural circuit function and translate findings into deep learning and neuromorphic circuits, enabling computing technologies inspired by nervous-system organizing principles and optimized for cognition. NeuronsXnets transfers knowledge through hands-on training across secondments, courses, workshops, and seminars, preparing a new generation of systems neuroscientists and computer scientists/engineers. Partners benefit via cross-fertilization, integrating results into existing and new solutions and strengthening long-term links among business, research, higher education, and hospitals, while mentoring students, raising public awareness, and pursuing research excellence in bio-inspired technologies.
Although response properties of individual neurons in primary visual cortex (V1) are well characterized, less is known about how functional ensembles—groups of neurons that co-activate more frequently than expected by chance—operate as computational units within V1 laminar microcircuits. Prior work identified pyramidal neuron co-activation patterns suggestive of ensemble organization, yet, despite increasingly detailed knowledge of structural connectivity, the functional principles governing ensemble structure and interactions remain unclear. We imaged pyramidal neurons across granular and supragranular layers of mouse V1 and applied pairwise functional connectivity analysis to identify multi-neuronal ensembles serving as putative information-processing modules. In the absence of visual stimulation, 19–34% of pyramidal neuron pairs within 300 μm were functionally connected, while 10% were connected at separations 1 mm. V1 laminae exhibited small-world organization, with layer-4 (L4) displaying slightly denser connectivity and a near-uniform degree-of-connectivity distribution. We propose that neurons together with their first-order functionally connected (1FC) partners constitute putative elementary units of cortical computation. The firing probability of layer 2/3 (L2/3) neurons depended on the number, not the identity, of co-active L4-1FC partners, with response sensitivity scaling to the aggregate distribution of L4-1FC inputs. Responses exhibited a ReLU-like nonlinearity emerging once 13% of L4-1FC partners co-fired, yielding sparse yet reliable responses. L2/3 neurons with L4-1FC modules of varying size showed distinct computational signatures and brain-state coupling. These properties were preserved under visual stimulation and different states of alertness. We have also applied this methodology on datasets collected in the context of passive learning, from mouse V1 L2/3 using 2-photon mesoscope imaging. This framework provides mechanistic insight into cortical circuit organization, while helping to link biological circuitry with deep-learning models of artificial intelligence.
We used mutual information to identify high predictive power (HPP) neurons for stimulus direction and characterize their functional connectivity. During stimulation, HPP neurons show higher event rates and disproportionately strong, dense connectivity that is largely distance-independent, consistent with a distributed, coordinated network; during rest, their event rates drop and their connectivity is indistinguishable from other responsive neurons, suggesting weaker shared drive from internal states. We also developed a two-phase PLSR method that decomposes activity into two orthogonal low-dimensional subspaces: a population subspace capturing shared global variability and a stimulus subspace that discriminates optical-flow directions while remaining linearly uncorrelated with the population subspace. Using mesoscopic two-photon calcium imaging in awake mice (L4, L2/3), we found that a small stimulus subspace preserves nearly all decoding performance, is stable over time, and is strongly correlated across mice. Removing the population component does not abolish stimulus discriminability, indicating direction coding does not depend on the global modulation.
Regarding the prefrontal cortex, neuronsXnets focused on how learning and cognitive training reshape circuit function and plasticity, complementing the project’s visual-cortex/network-analysis thrust. It complements visual-cortex analyses, by targeting cognitive circuits. A central contribution is the working-memory training pipeline, combining electrophysiology, ex vivo and in vivo calcium imaging (including miniscope recordings), and behavior to quantify how training alters synaptic and network properties. We found that working-memory training enhanced functional and structural plasticity in the prefrontal cortex and hippocampus. Training also reduced the frequency of spontaneous excitatory and inhibitory synaptic currents (sE/IPSCs) without altering the excitation/inhibition ratio, suggesting reduced low-frequency background activity and increased signal-to-noise ratio. Our on-going analysis of in vivo neuronal activity recordings using miniscope imaging will likely verify the above hypothesis. These datasets and conclusions directly support neuronsXnets’ goal of linking learning to circuit/network reconfiguration beyond sensory cortex. Another contribution is our finding that the human glutamate-metabolism gene glutamate dehydrogenase 2 (GLUD2) provides mice with enhanced learning and synaptic plasticity via modulation of astrocytic lactate metabolism. In parallel, the team characterized pre-ictal oscillatory signatures across hippocampal subregions in an ex vivo epileptiform model, highlighting predictive dynamics consistent with neuronsXnets’ interest in state-dependent network organization and transitions. Across these efforts, they enabled capacity-building through secondments and training (e.g. miniscope acquisition/analysis know-how and multi-terabyte datasets), strengthening the consortium’s experimental breadth and translational reach.
Finally, neuronsXnets developed a multimodal pipeline (fMRI, EEG, behavioral tasks) linking fluctuations in mental blanks and attentional to whole-brain network dynamics. It identified characteristic brain patterns and arousal-related factors, informing theories of spontaneous thought, consciousness & altered states.
Dissemination: 12 B.Sc. & 5 MSc thesis; 3 Ph.D. theses completed, 4 ongoing; 15+ interns; 9 postdocs; 18 Publications; 1 new course; several seminars/workshops
We used mutual information to identify high predictive power (HPP) neurons for stimulus direction and characterize their functional connectivity. During stimulation, HPP neurons show higher event rates and disproportionately strong, dense connectivity that is largely distance-independent, consistent with a distributed, coordinated network; during rest, their event rates drop and their connectivity is indistinguishable from other responsive neurons, suggesting weaker shared drive from internal states. We also developed a two-phase PLSR method that decomposes activity into two orthogonal low-dimensional subspaces: a population subspace capturing shared global variability and a stimulus subspace that discriminates optical-flow directions while remaining linearly uncorrelated with the population subspace. Using mesoscopic two-photon calcium imaging in awake mice (L4, L2/3), we found that a small stimulus subspace preserves nearly all decoding performance, is stable over time, and is strongly correlated across mice. Removing the population component does not abolish stimulus discriminability, indicating direction coding does not depend on the global modulation.
Regarding the prefrontal cortex, neuronsXnets focused on how learning and cognitive training reshape circuit function and plasticity, complementing the project’s visual-cortex/network-analysis thrust. It complements visual-cortex analyses, by targeting cognitive circuits. A central contribution is the working-memory training pipeline, combining electrophysiology, ex vivo and in vivo calcium imaging (including miniscope recordings), and behavior to quantify how training alters synaptic and network properties. We found that working-memory training enhanced functional and structural plasticity in the prefrontal cortex and hippocampus. Training also reduced the frequency of spontaneous excitatory and inhibitory synaptic currents (sE/IPSCs) without altering the excitation/inhibition ratio, suggesting reduced low-frequency background activity and increased signal-to-noise ratio. Our on-going analysis of in vivo neuronal activity recordings using miniscope imaging will likely verify the above hypothesis. These datasets and conclusions directly support neuronsXnets’ goal of linking learning to circuit/network reconfiguration beyond sensory cortex. Another contribution is our finding that the human glutamate-metabolism gene glutamate dehydrogenase 2 (GLUD2) provides mice with enhanced learning and synaptic plasticity via modulation of astrocytic lactate metabolism. In parallel, the team characterized pre-ictal oscillatory signatures across hippocampal subregions in an ex vivo epileptiform model, highlighting predictive dynamics consistent with neuronsXnets’ interest in state-dependent network organization and transitions. Across these efforts, they enabled capacity-building through secondments and training (e.g. miniscope acquisition/analysis know-how and multi-terabyte datasets), strengthening the consortium’s experimental breadth and translational reach.
Finally, neuronsXnets developed a multimodal pipeline (fMRI, EEG, behavioral tasks) linking fluctuations in mental blanks and attentional to whole-brain network dynamics. It identified characteristic brain patterns and arousal-related factors, informing theories of spontaneous thought, consciousness & altered states.
Dissemination: 12 B.Sc. & 5 MSc thesis; 3 Ph.D. theses completed, 4 ongoing; 15+ interns; 9 postdocs; 18 Publications; 1 new course; several seminars/workshops
NeuronsXnets’s cross-disciplinary synergies among neuroscientists, data scientists, imaging engineers, and systems developers provided ERs/ESRs with high-value multidisciplinary training, boosting their competitiveness as ML-driven neuroscience becomes central. Researchers gained hands-on experience with state-of-the-art tools (two-photon imaging, fMRI, SLM at Harvard; miniscopes at UCLA). The project also fostered soft and teaching skills, while networking and secondments expanded host-institution exposure, enabling new collaborations, broader professional networks, and stronger career prospects.