Two major recent advances have raised the possibility of fundamental breakthroughs in both basic and clinical neuroscience: the development of new tools to probe the nervous system with single-cell resolution as well as brain-wide scope, and breakthroughs in machine learning methods for handling complex data. Yet there remain crucial barriers to progress: while data acquisition tools are now broadly within the grasp of neuroscience researchers, the same cannot be said about data analytical tools that can tackle the complexities of the new data sets being gathered. In addition, the highly training-data dependent, black-box Artificial Neural Network (ANN) methods that have shown rapid growth in the technological domain, are not well-suited to scientific data analysis, where transparency and understanding is more important than black-box performance measures. This proposal brings together a cross-disciplinary team of leading neuroscience and computer science researchers to develop and deploy a critical set of data analytical tools for the neuroscience community. The tools will be useful for data already gathered in major group efforts in the US Brain Initiative, and also for new data sets being acquired using the tools developed in the Initiative. Extraction of the projection morphologies of individual neurons, and the classification and analysis of neuronal cell types is a central goal of the Brain Initiative. Because data from various sources are often analyzed with custom algorithms, scaling up existing approaches for use across large datasets and multiple data types has been a challenge. Instead researchers need a comprehensive, flexible mathematical framework that can be applied to a wide variety of data, including both static and dynamic measures. We propose to achieve this goal by combining Topological Data Analysis (TDA) methods with Deep Net based machine learning methods. Such a combined approach retains the flexibility of data-driven ANN methods while at the same time brings in conceptually well-grounded methods from TDA that are still able to address the complexities of brain-wide data sets with single-cell resolution.
Aim 1 of the proposal will use these methods to automate tasks in neuroanatomy previously requiring intensive human expert effort.
Aim 2 will apply the methods to single cell omics data sets.
Aim 3 will deploy the tools developed to the Brain Initiative Cell Census Network and the neuroscience community.
Recent years have seen many advances in experimental tools for probing brains in unprecedented ways, with single cell resolution, and as a result both individual investigators and large consortia are generating brain-wide single-cell data at an unprecedented scale. To derive full benefit from these data sets, researchers need theoretical and computational tools to analyze, visualize and derive knowledge from the data. In the proposed work, theoretically principled tools from Topological Data Analysis, in particular Discrete Morse Theory, are combined with artificial neural networks for Machine Learning, to provide a computational and analytical framework to deal with the complexity of the large-scale, high-dimensional data sets and derive the full benefits for basic as well as clinical neuroscience research.
