Program and Speakers

Until a few decades ago the basal ganglia were considered part of the motor system because of the conspicuous movement disorders associated with pathological changes in human Parkinson’s disease and Huntington’s disease. The discovery that dopamine projections from the midbrain to the striatum - the caudate-putamen and nucleus accumbens - mediated positive reinforcement and reward related learning led to recognition of striatal contributions to reinforcement learning. The mechanism appears to involve dopamine modulation of activity-dependent synaptic plasticity. This occurs at the synapses connecting the cerebral cortex to the striatum, where there is convergence of dopamine projections with cortical afferents to striatal output neurons. We have shown that these corticostriatal synapses exhibit synaptic plasticity that depends on three factors: presynaptic and postsynaptic activity, plus dopamine. The architecture of the striatum is essentially that of a single-layer, perceptron-like neural network as envisaged by Rosenblatt, and the biological synaptic plasticity rules are similar to the perceptron learning algorithms.  Insights from learning theories of Rescorla-Wagner, converged surprisingly with development of reinforcement learning algorithms by Barto and Sutton. Further on, we recognise that reinforcement learning enables adaptive responses to environmental contingencies, but in a rapidly changing environment it is important to behave even more flexibly. Recently we have been able to study the contibution of striatal cholinergic interneurons to behavioral flexibility, thanks to technical advances that allow visualization and manipulation of striatal acetylcholine. These studies have brought a fourth factor into focus, by which acetylcholine modulates dopamine-dependent synaptic plasticity. Understanding how this mechanism works in the striatal network might lead to better models of reinforcement learning, a better understanding of the symptoms of basal ganglia disorders, and might have implications for machine learning based on interactions with a changing environment.

Attention-deficit/hyperactivity disorder is a common neurodevelopmental condition characterized by developmentally inappropriate levels of inattention, hyperactivity and/or impulsivity. Altered dopaminergic functioning is proposed to contribute to the development of ADHD behaviors, through altered sensitivity to reinforcement and effects on learning. Individuals with ADHD are proposed to show impaired conditioning to reward predicting cues, and instrumental learning deficits are assumed to arise when reinforcement is delayed or non-continuous. Theoretically, these learning deficits have important implications for psychotherapeutic interventions for those with ADHD, which are based on reinforcement learning principles. Proposed neurobiological alterations underlying ADHD behaviors will be explained, followed by findings from empirical studies conducted in the Tripp unit, addressing theoretical predictions. Implications for psychotherapeutic interventions will be elaborated upon to highlight the translational value of experimental research.

Like human infants learn to speak, juvenile songbirds learn to sing by memorizing and then matching their vocalizations to their tutor’s song (TS) during the developmental song learning critical period. In zebra finches, a commonly used songbird model, only males learn to sing an individually unique song and keep it for their entire lives. Recently, we found transient sensorimotor projections into the song premotor area, HVC, from the neurons activated by hearing TS in zebra finch higher auditory area, caudomedial nidopallium (NCM). In contrast to the dense NCM axonal projections observed in HVC in the juveniles in sensorimotor song learning period, significantly lower amount of NCM-HVC projections were found in the older juveniles at the end of song learning period. This suggests that NCM-HVC projections disappear during song crystallization (Louder et al., 2024). Here, we further examined the dynamics of NCM projections within HVC using longitudinal structural two-photon in vivo imaging and confirmed the disappearance of NCM-HVC projections at the end of song learning period. We then investigated whether song experiences and their timing affect the time course of NCM-HVC projection dynamics and song learning. We raised the juveniles with two tutors sequentially and thus made juveniles learn to sing from the first tutor (T1), and then from the second tutor (T2) after an isolation period. We found that the sequentially tutored juveniles learned songs from T2 beyond the normal learning period. Also, their songs were less stable than songs of normally reared birds at the end of normal sensorimotor learning period. We further visualized the NCM neurons responsive to each tutor’s song in sequentially tutored birds by using a combination of viruses which express fluorescent proteins in an activity-dependent manner (AAV9-cFos-TetON-EGFP and AAV9-cFos-TetOFF-mRFP). The projections from NCM neurons responding to T2 song were found in HVC within 12 hrs after being exposed to the T2 in sequentially tutored juveniles. Notably, projections remained in HVC from both NCM neurons activated by hearing T1 or T2 songs after sequential learning from two tutors in adulthood when NCM-HVC projections were rarely found in normally raised birds. Taken together, our results suggest that early song learning experiences and their timing shape auditory-motor neuronal circuits and learning time course.

The initiation of skilled movements is regulated by descending motor pathways projecting to the spinal cord. However, the mechanisms by which these pathways convey initiation signals remain unclear. This uncertainty arises from two key published observations. First, inhibition of each descending pathway does not abolish the initiation of skilled movements such as reach–grasp behaviors. Second, brief stimulation of descending neurons fails to evoke the sequence of these movements. In this study, we demonstrate that combined inhibition of corticospinal neurons (CSNs) in the rostral forelimb area (rCSNs) and reticulospinal neurons (ReSNs) in the reticular formation effectively blocks the initiation of skilled reaching behaviors. Conversely, simultaneous brief activation of both rCSNs and ReSNs is sufficient to not only initiate reaching, but also to evoke sequential, naturalistic reach–grasp movements. These findings suggest that rCSNs and ReSNs converge on a “reach–grasp generator” located in the cervical spinal cord. Supporting this, we find that both rCSNs and ReSNs are activated prior to movement initiation, and that these two descending circuits operate independently, with no apparent crosstalk in the central nervous system. Together, these findings reveal a cooperative mechanism by which two parallel descending pathways drive the initiation and execution of skilled motor sequences in mammals.

Neurons are highly polarized cells with an elongated axon that extends far away from the cell body. In order to maintain neuronal homeostasis, neurons rely extensively on axonal transport of membranous organelles and other molecular complexes in addition to local translation of proteins. Axonal transport plays a central role in the establishment of neuronal polarity, axonal growth and stabilization and synapses formation, allowing for precise spatio-temporal activation and modulation of numerous molecular cascades. Anterograde and retrograde axonal transport is supported by various molecular motors, such as kinesins and dyneins, and a complex microtubule network. In this seminar I will discuss some aspects of retrograde signaling in neurons, ranging from injury signals to dynein-mediated axonal transport, which are critical for the survival of neurons. We will also discuss strategies to promote axonal regeneration through to use of specialized substrates and the tools we have developed to investigate the mechanisms underlying axonal degeneration in Amyothrophic Lateral Sclerosis (ALS).

Vertebrates and cephalopods have independently developed camera-type eyes — a striking example of convergent evolution. While many vertebrates detect color through wavelength-sensitive retinas, cephalopods have instead evolved retinas that detect light polarization. Polarization vision is thought to confer ecological benefits in underwater environments, such as enhanced object detection and intraspecies communication. The cephalopod visual center, the optic lobe (OL), exhibits a highly organized and distinctive structure with two parts: the cortex, which bears morphological similarity to the vertebrate retina (hence often termed the “deep retina”), and the medulla, which has a characteristic tree-like anatomical organization. However, the mechanisms by which the OL integrates luminance and polarization signals to support cephalopod behavior remains elusive, largely because in vivo neural recordings have been challenging to implement in these soft-bodied animals. Here, we developed a novel head-fixation method to perform two-photon in vivo calcium imaging in the brain of awake juvenile squids (S. lessoniana). We recorded calcium responses from populations of OL cortex neurons while delivering visual stimuli that varied in intensity and polarization. Our data revealed distinct neuronal classes defined by spatiotemporal tuning and intensity–polarization specificity. These included (i) integrators of horizontal and vertical photoreceptor inputs, (ii) neurons selective for a single polarization orientation, and (iii) neurons that compute the degree of linear polarization (DoLP) by subtracting orthogonal signals. We also observed orientation- and direction-selective neurons in the inner granule layer, but rarely in the outer granule layer, suggesting a feed-forward architecture for progressively complex feature extraction. We then performed in vivo electrophysiological recordings using Neuropixels probes, targeting downstream regions of the OL cortex including the medulla and peduncle lobe. We found that receptive field sizes increased with OL depth, a hallmark of hierarchical visual processing. Medulla neurons often displayed additive integration of intensity and polarization contrasts, while others showed strongly nonlinear selectivity. Anatomical experiments, including single-neuron dye filling and brain-wide retrograde tracing, further supported a model of hierarchical visual information processing through the medulla’s tree-like organization. Collectively, our study provides the first in vivo physiological characterization of the cephalopod visual system, offering new insights into the neural mechanisms underlying their specialized underwater vision as well as broader features of convergent evolution in visual systems.

Coleoid cephalopods (squids, octopus, cuttlefish), with their complex visual system and exhibiting sophisticated visually-guided behaviour, have in recent years become a popular model system in neuroscience. However, little is known about the cells (and their connections) comprising the optic lobe (OL) - the main processing centre for visual information. Here, we set out to characterise and classify the cells of the OL of the bigfin reef squid Sepioteuthis lessoniana. To build a comprehensive cell atlas, we plan to measure the (a) morphological, (b) electrical, (c) biochemical, (d) genetic, and (e) functional cell properties, and most importantly, establish a link between them. We have developed an acute slice preparation of the optic lobe which allows us to combine electrophysiological recordings, morphological reconstructions, pharmacological intervention, and single-cell RNA sequencing. I will present our initial exploration of morphological cell classes of the OL cortex and ongoing work linking these to biochemical, electrophysiological, and functional properties.

Motion can play a major role in bioimaging, either as an artifact or as a physiological indicator. I will present Flow-Registration, an optical flow-based motion correction approach that is tailored to the statistics of 2-photon microscopy videos. Flow-registration can describe and correct large, non-uniform motion with sub-pixel precision, amplify small, non-uniform motion, reconstruct 3D stacks, estimate z-drift and is ready for the next generation of 3D imaging approaches. Our method produces state-of-the-art results on low-SNR data with a wide range of motion artifacts. Flow-Registration is available as a MATLAB toolbox, a Python package, and plugins for Fiji and napari, enabling both batch pipelines on large data and explorative data analysis (https://github.com/FlowRegSuite). I will present how the toolbox can be integrated into various optical neuroimaging workflows and give some examples.

All-optical strategies enable to identify functional neuronal ensembles with calcium imaging and replay/alter their spatio-temporal activity with optogenetics to decipher their behavioral implications. We have previously developed the first fiber-coupled microscope enabling two-photon (2P) functional imaging and 2P holographic photostimulation with near-single cell resolution in freely moving mice, named 2P-FENDO. Here, we present a significantly optimized 2P-FENDO-II system that achieves a four times larger field of view, a more homogeneous light distribution across the field of view, both for imaging and photostimulation, while achieving better flexibility and thus optimal adaptation to the study of freely moving mice. We have demonstrated the performance and versatility of 2P-FENDO-II in experiments targeting the somatosensory cortex, the visual cortex or the cerebellar cortex, in which we showed concomitant calcium imaging with jGCaMP7s and optogenetic control with ChRmine. These enhancements establish 2P-FENDO-II as a groundbreaking tool for investigating on large volume complex neuronal dynamics and behavior with unprecedented detail in naturalistic situations.

I will give an overview of the current projects in the Optical Neuroimaging Unit. I will quickly introduce voltage two-photon imaging, Ca2+ imaging in cerebral and cerebellar cortex in neurons and astrocytes, PKA imaging, expansion microscopy, and more to inspire discussions.

Voltage-gated ion channels are crucial for signal processing and propagation in neurons. Many of these channels are selectively permeable to sodium, potassium, or calcium, with each type playing a distinct role in neural signaling. Modeling the voltage-dependent gating of ion channels has a long history, dating back to Hodgkin and Huxley (1952). However, recent studies have revealed that genetically identical channels expressed in the same cell type can exhibit significant kinetic heterogeneity. Classical Hodgkin-Huxley-like (HH-like) models are not designed to capture this heterogeneity, making it difficult to model and limiting our understanding of channel behavior. Additionally, building HH-like models is typically complex and time-consuming due to the need to choose functional forms, estimate numerous parameters, and make multiple modeling decisions. In this talk, I discuss the application of two approaches -- non-linear mixed effects modeling (NLME) and scientific machine learning (SciML) -- that address the challenge of modeling kinetically heterogeneous gating. SciML combines machine learning techniques with mechanistic scientific models, while NLME accounts for both within- and between-entity variability. By integrating these approaches, we constructed a unified SciML HH-like model capable of fitting recordings from 20 different Kv channel types at three temperatures. The unified model achieved significantly better fits than seven previously published HH-like models. Therefore, NLME and SciML may be valuable tools in understanding the voltage-gated ion channel gating and other sources of heterogeneous neural data.

Bayesian inference is a standard way of handling uncertainties in sensory perception and reinforcement learning is a common way of acting in unknown environments. While they are used in combination for perception and action in uncertain environments, the similarity of their computations has been formulated as the duality of inference and control, or control as inference. Meanwhile, an intriguing question about the brain is why the entire neocortex shares a canonical six-layer architecture, while its posterior and anterior halves are engaged in sensory processing and motor control, respectively. Here we consider the hypothesis that the sensory and motor cortical circuits implement the dual computations for Bayesian inference and optimal control, and report the ongoing neural recording experiments to test the hypothesis.

Can a robot learn to adapt its goal-directed actions in a dynamic environment by learning from a human's examples? In this work I demonstrated a real-time active inference model that learns incrementally from a few human tutoring episodes, similar to how a caregiver might physically guide a child, and from its own mental rehearsal. By framing goal-directed action as inferring latent trajectories that minimize expected free energy, robust planning without exhaustive policy search is achieved. In physical robot experiments, a humanoid robot learned to adapt when objects or goals suddenly moved after only a handful of demonstrations, while avoiding catastrophic forgetting of previously learned skills.

Rett syndrome is a development neuronal disorder and one of the most common causes of mental retardation in women. Sporadic mutations on the transcription factor MeCP2 has been found to be the most frequent cause for Rett syndrome in humans, but despite numerous studies, the mechanism for the causation of Rett syndrome by MeCP2 mutations is still largely unknown (Good et al., 2021). Several animal models of Rett syndrome had been developed in mice and zebrafish (Pietri et al., 2013), but until now there were no descriptions of how MeCP2 mutations affect adult zebrafish behavior. Here, I characterized the behavioral phenotype of these MeCP2 mutants in particular fear and anxiety behaviors using several well stablished behavioral tests. MeCP2 mutants show drastically elevated levels of anxiety and increased fear response, a phenotype reminiscent of other animal models and human patients of Rett syndrome. To study what neural circuits and brain areas are responsible for these changes in behavior, I compared the levels of several brain metabolites, observing a significant reduction of epinephrine and norepinephrine levels. Imaging analysis of the brain, and in particular of the telencephalon, using a transgenic line expressing GFP in the Dm (Lal et al., 2018), showed an increase in size of the Dm region of the dorsal telencephalon, the functional equivalent to the basolateral amygdala in mammals and region responsible for anxiety and fear behavior. Finally, I visualized c-Fos activation in several brain regions after Alarm substance treatment observing a strong increase neural activation on the Dm and other brain regions. Suggesting that MeCP2 mutation induced defects on the anatomy of the telencephalon are likely involved in the behavioral deficits of this mutants regarding fear response.

Abstract will be pasted here.

We have developed a new algorithm, based on dynamical systems theory, that can map neural activity time series to corresponding behaviors—provided that the behavioral information is present and the signal-to-noise ratio is sufficient. The method is a greedy implementation of Takens’ theorem, selecting subsets of time series from a larger library to reconstruct attractor dynamics and predict a target behavioral variable with optimal accuracy. Importantly, the algorithm identifies specific neural observables that contribute to each behavior, making its predictions experimentally testable. We demonstrate its versatility using diverse datasets, including lightsheet calcium imaging in Drosophila, whole-brain calcium imaging in zebrafish larvae, tetrode recordings in rats, fMRI data in humans, and Neuropixels recordings in mice. A user-friendly web application will soon be available for the OIST community to apply this method to their own datasets.