C&SP27: Circuit reconstruction of association cortex

A. Collaborating Investigators: Wei-Chung Allen Lee,1 Art Wetzel,2 Greg Hood2

B. Institutions: 1Harvard Medical School (HMS) and 2Pittsburgh Supercomputing Center

C. Funding Status: NIDCD 5R01-DC013622-03" Network Anatomy of Olfactory Processing (Lee) 8/9/2013 - 8/31/2016; DP2 OD022472-01 'Network anatomy of behavioral choice' (PI: Lee) Pending

D. Biomedical Research Problem:

This C&SP with Wei Chung Allen Lee at Harvard Medical School builds upon the successful DBP5 with Clay Reid of Harvard and the Allen Brain Institute during the 2012-2017 funding period. Lee is a former member of Reid's lab, and has continued this line of connectomics research as an independent researcher at HMS using both the original TEMCA microscope,87 and an improved successor, TEMCA-GT. Our previous collaborative work87,88with Reid and Lee has focused on studying the pattern of anatomical connections among functionally characterized neurons within mouse visual cortex. Lee plans to study next a new cortical area in the mouse. This region of the cortex is hypothesized to be involved in key cognitive functions including decision-making, movement planning, attention, and reward.89-93 Recent work has demonstrated that groups of neurons in this brain region are activated sequentially.94 Neurons fire selectively during discrete epochs of a navigational task, and the resulting sequences of neuronal activity accurately predict the behavioral choice of the animal.94 Two of the unique aspects of this project are that we will, for the first time, be able to compare an association area with a sensory area and identify conserved and distinguishing features of connectivity motifs underlying their different network dynamics.



Fig VIII.7. Preliminary acquisition of a10 gigapixel section from a cortical area of mouse brain involved in decision-making investigated by the Lee lab.


E. Methods and Procedures: We will trace a portion of the neural connectome within a region of tissue about 1mm3 in size. To do this will require ~1Petabyte of raw image data comprising ~25,000 sections. This volume will encompass 4 cortical columns, which are hypothesized to be modular, canonical, local circuits. Prior datasets have been up to Terabyte in size, so this will necessitate an order-of-magnitude improvement in our ability to acquire EM images and analyze them. Several technical hurdles must be overcome, and we focus here on the issues affecting alignment of these raw images into a form which can then be traced, either manually or in a semi-automated way. Specifically, we must reduce the amount of manual intervention so that the effort is much less per gigapixel than on prior datasets. Most of the manual intervention that has been necessary in aligning prior datasets has been associated with artifacts introduced during sectioning and handling (e.g., tissue folds and tears), so an important step in reducing labor has been the development of an improved automated pipeline for sample handling. This new workflow is centered around a novel substrate that allows automated collection of 1000s of serial thin sections from the microtome, and imaging them in the vacuum chamber of the TEMCA-GT transmission electron microscope. TEMCA-GT's improved stage allows for more accurate positioning of the samples, and this information can be fed into our AlignTK pipeline to constrain the model’s initial conditions and reduce the probability of registration errors (Fig VIII.7). This new approach will allow 10 times more data (5-10 TB) to be imaged per day, resulting in an increased workload for the alignment software. AlignTK is being converted to use GPUs for its most computationally-intensive tasks (TR&D2 Subaim 2.1). With current Harvard Medical School and MMBioS facilities, we estimate we will be able to register 64 Mpixels/sec of image data, with the ability to process a ~1PB dataset in roughly 6 months. The main concern will be to increase the robustness of the software to deal well with atypical image content (e.g. voids within capillaries), and to make it easy to incrementally process the data as it is generated. To increase robustness and throughput, we may also perform registration using MMBioS' SWiFT-IR software, which has heretofore been used primarily on SEM images, but should also perform well with TEM data.


Copyright © 2018 National Center for Multiscale Modeling of Biological Systems. All Rights Reserved.