A scalable cloud-based framework for multi-modal mapping across single neuron omics, morphology and electrophysiology

Broad Institute, Inc., Cambridge MA

Investigators

Abstract

Project Summary Categorizing individual neurons into different groups, or cell types, is a classical approach to studying the nervous system. With increasingly more tools being invented to observe the neurons, new criteria were created to characterize different aspects, or modalities, of the cells. While these modality-specific categorizations have enabled in-depth knowledge in neuroscience, the inconsistencies across different criteria leave data integration across modalities technically difficult. In addition, disagreements of “similar cells” using different categorization criteria have resulted in division of the neuroscience community into modality-centric subgroups. To solve this problem, objective approaches to define cell similarities incorporating multiple modalities must be established and developed in a way that is open, accessible, and engaging to the neuroscience community at large. We propose to develop a broadly accessible cloud-based framework toward an integrative, multi-modal brain cell atlas using novel, scalable analytics tools, leveraging federated BRAIN Initiative resources and community engagement. The self-supervised learning methodology is purely data-driven, and allows highly accurate identification of similar cells based on single or multiple modalities of measurements, without presumption about modality-specific classes. It also enables cross-modality mapping, where unobserved measurements can be inferred from single-modal data, achieving computational “scale-up” of the brain cell mapping efforts in low- throughput modalities such as electrophysiology or morphology. With more multimodal data being collected and added to the study, the algorithm accuracy of this inference will continue to grow in an automated way. The open-source methodology will be productionized into cloud-native pipelines for individual modalities and for cross-modality mapping, and installed in a cloud computing workbench as a part of the ecosystem, for scalable, open data analyses. This cloud ecosystem will demonstrate access to BRAIN Initiative data repositories hosting molecular (NeMO), neurophysiology (DANDI), and microscopy (BIL) data, such that datasets from different sources can be brought into a common workspace for integrative analyses. Furthermore, the cloud ecosystem will provide a user-friendly data portal for visualization and navigation of the multi-modal single cell data from the repositories, as well as exploring the cell similarity query results. This ecosystem will support FAIR principles and promote collaborative research and seek for extended integrations with data repositories. Through broad engagement and outreach to neuroscience communities, this project will provide resources for building an integrated brain cell atlas and facilitate the multimodal characterization of the brain.