Data driven mapping of the drosophila larval central nervous system

Abstract

The Central Nervous System (CNS) of the larval Drosophila model organism is extensively studied partly due to its small size and short generation times but also due to its ability to learn and the availability of genetic tools to investigate individual cell function. Unfortunately, it is very difficult to pool data from different studies: There is a lack of a standardised reference atlas and inference among separate 3D image stacks from different individual larvae is slow and error-prone. If, however, identical cells from images of different genetic lines can be found, this cell type can be isolated and probed for function via the Split-GAL4 method. The principal aims of my work were to find, implement and test methods that can be used to automate this process and analyse combined cell imaging data for information about the gross neuroanatomy of the larva. I annotated a template larval Central Nervous System with neuropile domains and lineage tracts from the literature and compiled the most complete textual domain descriptions to date for the FlyBase database. To develop a registration pipeline for the whole-CNS channel of over 22 000 image stacks with a signal channel sparsely populated with neurons, I evaluated non-rigid registration parameters by measuring overlap of registered identical neurons. B-Spline Free-Form Deformations with a Correlation Ratio similarity metric were performed and candidate cell volumes extracted using adaptive thresholding. I evaluated registration accuracy with a novel local-intensity difference algorithm implemented with dynamic programming, yielding over 6 500 satisfactory individual whole-cell images. I applied Machine Learning to identify neuron somas in semi-automatic cell annotation. To find similar neurons, I implemented and evaluated the established nBLAST method and developed a new approach: This condenses the representation of neurons with computer vision Artificial Intelligence (Convolutional Neural Networks within a triplet network architecture). These methods successfully allow biologists to rank cells by similarity, with the novel method demonstrating similar accuracy but executing 30 times faster. I validated this new method further by hierarchical clustering of cell examples to attempt to find cell type clusters. To create an average representation of a cell type from many examples, I developed a novel algorithm. Finally, I have shown that voxel clustering on cell expression patterns supports the existence of most larval neuropil domains, with the notable exception of the Clamp. The registered cell examples have been made available as part of the freely accessible and actively used larval Virtual Fly Brain atlas.