Periodic Reporting for period 1 - ETIC (Encoding and Transmission of Information in the Mouse Somatosensory Cortex)
Berichtszeitraum: 2016-07-01 bis 2018-06-30
"Brain diseases are taking an increasing toll on aging European Societies, and being able to cure them and limit the costs and the social exclusion they generate is advocated by many European policies. Yet, due to its enormous complexity, the brain is the least understood organ of all. Importantly, in order to be able to treat a malfunctioning brain we first need to understand how a normal brain works, which requires that we first crack the neural code, i.e. the language neurons use to encode and transfer the information they receive from the external world.
To address this question much effort has been devoted to record the responses of cortical sensory areas to experimentally controlled stimuli and to build a “dictionary” that can be used to understand how an external event is encoded in the activity of a neural population. Although this so-called “Rosetta Stone” approach has provided much knowledge about the neural code in the past decades, it suffers from three important limitations. First, neurons are noisy: the very same external stimulus can elicit different responses on a neural population, which makes difficult to tear apart what is noise from what constitutes relevant information. Second, to fully describe neural population activities an enormous number of variables –that increases exponentially with the number of neurons- are needed. Third, the ""Rosetta Stone"" approach does not address the question of whether a putative neural code that carries some sensory information is then transmitted to downstream networks.
For these reasons, we need to complement statistical approaches like the “Rosetta Stone” with more direct, causal, techniques that allow manipulating the activity of specific sub-populations of neurons. Recently developed optogenetics approaches allow controlling the neuronal activity using light-gated proteins, and therefore provide a way of testing the role of a specific encoding strategy by applying appropriate stimulation protocols to a population of neurons and evaluating the effects of such stimulation, for instance on a post-synaptic network. However, to be correctly applied to neuroscience experiments these techniques need to be complemented by novel theoretical developments that: 1) identify hypotheses about the encoding strategies used by the neural population under study; 2) create stimulation protocols specifically designed to test the identified hypothesis."
To address this question much effort has been devoted to record the responses of cortical sensory areas to experimentally controlled stimuli and to build a “dictionary” that can be used to understand how an external event is encoded in the activity of a neural population. Although this so-called “Rosetta Stone” approach has provided much knowledge about the neural code in the past decades, it suffers from three important limitations. First, neurons are noisy: the very same external stimulus can elicit different responses on a neural population, which makes difficult to tear apart what is noise from what constitutes relevant information. Second, to fully describe neural population activities an enormous number of variables –that increases exponentially with the number of neurons- are needed. Third, the ""Rosetta Stone"" approach does not address the question of whether a putative neural code that carries some sensory information is then transmitted to downstream networks.
For these reasons, we need to complement statistical approaches like the “Rosetta Stone” with more direct, causal, techniques that allow manipulating the activity of specific sub-populations of neurons. Recently developed optogenetics approaches allow controlling the neuronal activity using light-gated proteins, and therefore provide a way of testing the role of a specific encoding strategy by applying appropriate stimulation protocols to a population of neurons and evaluating the effects of such stimulation, for instance on a post-synaptic network. However, to be correctly applied to neuroscience experiments these techniques need to be complemented by novel theoretical developments that: 1) identify hypotheses about the encoding strategies used by the neural population under study; 2) create stimulation protocols specifically designed to test the identified hypothesis."
How to mathematically separate out the different components of a neural code and to identify the unique contribution of each of these components to sensory coding and behavior is an open question in neuroscience. Answering this question will be an important step towards understanding the way neurons process the information they receive. During the ETIC project, I investigated in detail the information encoded in the spiking activity of a neuron by computing the unique contribution that each temporal scale makes to it. I have done this by analytically inferring the derivative of the information with respect to the precision with which the neural response is measured. The Information Jitter Derivative (IJD) method, as we called it, allows inferring the temporal scales playing a relevant role in the encoding of the information about a given set of stimuli (see Fig. 1).
In a second step, in collaboration with Tommaso Fellin’s laboratory, I developed a novel protocol that permits applying two-photon functional imaging at high sampling rates, which increases the signal to noise ratio of the collected fluorescence traces (see Fig. 2). The Smart-line scan method is based on the careful design of scan trajectories that only sample from areas of interest (in our experiments these are pyramidal cells in layer 4 of the somatosensory cortex of the mouse expressing the GCaMP6s calcium indicator). I further confirmed the results obtained in the real experiments by developing experimentally-informed simulations that allowed investigating the smart-line scan technique in several realistic scenarios.
Finally, I extended the Generative Adversarial Networks (GANs) framework to model the activity of the population of neurons under study. Spike-GAN, as we called our method, allows producing highly realistic spike trains matching the first and second-order statistics of a population of neurons. More importantly, we can use the trained network to produce what we called importance maps that allow detecting, in a given activity pattern, those spikes participating in a specific feature characteristic of the probability distribution underlying the training dataset (see Fig. 3). Therefore, Spike-GAN constitutes a powerful tool for unsupervised identification of highly salient low-dimensional representations of neural activity, which can then be used to elaborate hypotheses about the encoding strategies used by a given neural population and discover the key units of neural information used for functions such as sensation and behavior.
In a second step, in collaboration with Tommaso Fellin’s laboratory, I developed a novel protocol that permits applying two-photon functional imaging at high sampling rates, which increases the signal to noise ratio of the collected fluorescence traces (see Fig. 2). The Smart-line scan method is based on the careful design of scan trajectories that only sample from areas of interest (in our experiments these are pyramidal cells in layer 4 of the somatosensory cortex of the mouse expressing the GCaMP6s calcium indicator). I further confirmed the results obtained in the real experiments by developing experimentally-informed simulations that allowed investigating the smart-line scan technique in several realistic scenarios.
Finally, I extended the Generative Adversarial Networks (GANs) framework to model the activity of the population of neurons under study. Spike-GAN, as we called our method, allows producing highly realistic spike trains matching the first and second-order statistics of a population of neurons. More importantly, we can use the trained network to produce what we called importance maps that allow detecting, in a given activity pattern, those spikes participating in a specific feature characteristic of the probability distribution underlying the training dataset (see Fig. 3). Therefore, Spike-GAN constitutes a powerful tool for unsupervised identification of highly salient low-dimensional representations of neural activity, which can then be used to elaborate hypotheses about the encoding strategies used by a given neural population and discover the key units of neural information used for functions such as sensation and behavior.
The Information Jitter Derivative (IJD) method constitutes a completely novel approach for quantifying the information that a particular temporal scale adds to coarser temporal scales. I demonstrated the usefulness of the IJD method on real neural responses recorded from the retinal ganglion cells in the salamander retina and from the trigeminal ganglion cells in the rat somatosensory system, showing that both populations of neurons carry the information about fine and coarse features using different temporal scales.
The Smart-line scan technique has the potential to be an instrumental technique allowing for the exploitation of the high sensitivity of the GCaMP6 calcium indicators to action potentials. We confirmed this potential both in real and in-silico experiments.
In contrast with previous methods, Spike-GAN reproduces the spatio-temporal statistics of neural activity without the need for these statistics to be handcrafted in advance, which avoids making a priori assumptions about which features of the external world make neurons fire. A promising application of Spike-GAN is that of designing realistic patterns of stimulation that can be used to optically perturb (using optogenetics) populations of neurons in order to test hypotheses about their encoding strategies.
The ETIC project had as its main goal the development of a novel theoretical and computational framework to investigate how the mammalian cortex encodes sensory information, using the mouse somatosensory cortex and layer 4 networks as an experimental model. Crucially, the techniques developed during the project will be generally applicable to any brain network and will thus lay the foundation for the progress on cracking the neural code in different structures of the mammalian brain. The ability to first “read” out (using Smart-line scan technique) and crack a neural code (using the IJD and Spike-GAN methodologies) from brain recordings and to then “write” its information on the neural tissue with optogenetics is absolutely central to the further development of brain machine interfaces that can for example write down sensory information into a cortical area not receiving inputs from the periphery because of injury or illness. Such brain-machine interfaces have a great potential to restore communication, health and social inclusion to a wide population of patients.
The Smart-line scan technique has the potential to be an instrumental technique allowing for the exploitation of the high sensitivity of the GCaMP6 calcium indicators to action potentials. We confirmed this potential both in real and in-silico experiments.
In contrast with previous methods, Spike-GAN reproduces the spatio-temporal statistics of neural activity without the need for these statistics to be handcrafted in advance, which avoids making a priori assumptions about which features of the external world make neurons fire. A promising application of Spike-GAN is that of designing realistic patterns of stimulation that can be used to optically perturb (using optogenetics) populations of neurons in order to test hypotheses about their encoding strategies.
The ETIC project had as its main goal the development of a novel theoretical and computational framework to investigate how the mammalian cortex encodes sensory information, using the mouse somatosensory cortex and layer 4 networks as an experimental model. Crucially, the techniques developed during the project will be generally applicable to any brain network and will thus lay the foundation for the progress on cracking the neural code in different structures of the mammalian brain. The ability to first “read” out (using Smart-line scan technique) and crack a neural code (using the IJD and Spike-GAN methodologies) from brain recordings and to then “write” its information on the neural tissue with optogenetics is absolutely central to the further development of brain machine interfaces that can for example write down sensory information into a cortical area not receiving inputs from the periphery because of injury or illness. Such brain-machine interfaces have a great potential to restore communication, health and social inclusion to a wide population of patients.