Periodic Reporting for period 1 - DeepNeuronEmbeddings (Deep Neuron Embeddings: Data-driven multi-modal discovery of cell types in the neocortex)
Berichtszeitraum: 2022-06-01 bis 2024-11-30
Understanding the relationship between structure and function of cortical neurons and circuits is one of the key challenges in neuroscience. For inhibitory neurons, roughly 15 subtypes are well characterized and we know a fair bit about their function. However, the vast majority of neocortical neurons are excitatory. Yet we know little about how differences in the morphology of excitatory neurons relate to their computational properties in vivo. In this project we hypothesize that there is a close correspondence between morphology and function of excitatory neurons: distinct subtypes can be identified not only by their morphological features, but also by how they respond to stimulation with natural stimuli. To test this hypothesis, we build upon recent advances in machine learning and develop a data-driven approach to derive a “bar code” for each neuron: a low-dimensional representation of its morphological features and its response properties to natural stimuli. Using these techniques, we tackle the structure-function question by harnessing a large-scale functional anatomy dataset: a combination of electron-microscopy reconstructions at sub-micrometer resolution with two-photon functional imaging of nearly all excitatory neurons in one cubic millimeter of the mouse visual cortex. If successful, our project could fundamentally change our view on the diversity of excitatory cell types and reveal how morphological features are linked to a neuron’s computational output. It could pave the way towards a unified definition of cell types, one of the fundamental building blocks of the brain. The same approach could be used in other brain areas and even other cellular systems beyond the brain. More broadly, while machine learning is promising to transform the scientific discovery process as a whole, our project could serve as a prime example of this transformation process in neuroscience and show how machine learning can help to discover structure in nature.
We focused on two main objectives. First, we successfully developed a method to analyze the 3D shapes of neurons, finding that neuron structures in the mouse visual cortex form a continuum rather than distinct types. Second, we made progress in predicting how neurons respond to natural visual stimuli and explored new ways to classify neuron functions, including a method that groups neurons based on their response to specific stimuli. Third, we began linking neuron shapes to their functions, though this work is still ongoing.
We achieved two important results: First, we developed a new method to analyze the shapes of neurons without needing predefined characteristics from experts. This method uses the raw structure of neurons to create "morphological bar codes." With this approach, we found that the structure of neurons is better understood as a continuous spectrum rather than distinct types, and we discovered a unique feature in certain brain cells that had not been known before. Second, we made significant progress in understanding how visual neurons function by improving predictive models and methods for grouping neurons based on their behavior.