From the smallest fragment of brain tissue, the intricate blueprint of the entire brain is beginning to emerge. Researchers at Baylor College of Medicine are making several time-consuming aspects of this process a lot easier with the development of a software package called NEURD short for "NEURal Decomposition". This new software increases the speed of data error detection and correction for the wiring, or "street map," of connections between cells in the brain, enabling new discoveries.
Published in the current edition of Nature, researchers describe how NEURD simplifies the process of preparing intricate datasets for a variety of downstream analyses - it proofreads, cleans up and provides an easier way to query and annotate the data with interpretable features. Furthermore, the authors show how using NEURD, this cleaned and annotated data can be easily explored, highlighting several new findings about the brain.
"The primary goal of this collaboration has been to uncover the importance of wiring in the brain, to reveal the biological 'secret sauce' that enables the brain's extraordinary problem-solving capabilities, which remain beyond the reach of current machine learning systems," said Dr. Jacob Reimer, assistant professor of neuroscience at Baylor, a founding member of the Center for Neuroscience and AI, and senior author on the study.
MICrONS
The collaboration Reimer is referring to is The MICrONS Project, and the associated dataset of millimeter-scale electron microscopy dataset that offers an unprecedented look at the wiring and function of the mouse brain. The project spanned seven years and involved more than 150 scientists from leading institutions around the world. (Read more about this project and the ten connected studies published simultaneously in the Nature family of journals.)
The global team created the largest, most detailed wiring diagram of a mammalian brain to date from a tiny slice of mouse visual cortex, 1 cubic millimeter in size. It contains more than 523 million synapses, miles of axons and more than 200,000 cells, and what makes this wiring diagram and brain map unique, is that it also includes functional properties of neurons (their electrical signals) as well as their structural and connectivity data, providing an unprecedented look at how neurons communicate and process information as a network - providing a better picture of the hidden conversations in the brain.
The MICrONS data not only includes high-resolution anatomical images, but it also integrates live functional data collected from the same cells. A small number of previous datasets have also matched function and structure, however, on a much smaller scale.
A unique imaging process was used to capture this activity. A team of researchers that Reimer was a part of as a graduate student in the Tolias Lab at Baylor (Tolias Lab currently is with Stanford University), spent weeks recording the responses of these neurons as the mice were exposed to different visual stimuli, such as movies like "Mad Max."
Researchers at the Allen Institute for Brain Science worked on the anatomical electron microscopy data collection from the extracted tissue, and collaborators at Princeton were responsible for the reconstruction of the massive (petabyte-scale) electron microscopy anatomical volume. That meant aligning all the images and segmenting them to create a 3D image of all the neurons and associated wiring.
This approach allows scientists to study both the structure and activity of the brain in tandem, providing an unparalleled view of how neurons in the brain process sensory information. Now with his own lab at Baylor, Reimer and the Tolias lab along with other collaborating institutions, have been working to try to understand principles relating function and structure. (Read more about functional connectomics in the mouse visual cortex, a part of the Nature package related to the MICrONS Project.)
Enter NEURD
To further refine the data and mine it for insights, the team at Baylor along with collaborators, developed NEURD, an advanced computational tool to proofread and clean the raw data, ensuring its accuracy for scientific use.
NEURD automates the proofreading process, identifying and correcting errors in the neural networks, "cleaning" up and annotating the 3D structure. Researchers performed extensive validation of the automated proofreading approach to determine the precision and recall of error correction, ensuring that subsequent analyses can proceed with as much accuracy as possible.
"NEURD tackles the challenge of automatic neuron proofreading by breaking it down into a sequence of simpler tasks that mirror how a human typically perceives a neuron. For instance, by first identifying features like spines, we can more easily infer cell type, which in turn helps distinguish axons from dendrites-ultimately making the proofreading process more constrained and manageable," said Dr. Brendan Celii, lead author on the study and formerly with Baylor and Rice University and currently with Johns Hopkins University Applied Physics Laboratory.
Building on existing software tools that create 3D mesh images of neurons, NEURD works through automatic detection of important structures like neuron cell bodies (soma), tiny protrusions on neurons called spines, and detailed segmentation of the neuron's axon and dendritic branches. These features are designed to work across a wide range of datasets, helping researchers gather consistent data for different types of experiments.
"Proofreading is only one aspect of NEURD," Reimer said. "The other aspect is to conduct science."
As more detailed maps of neural connections are created, spanning various species and brain regions, NEURD will be crucial for advancing the understanding of neural networks. While this research is still in its early stages, the implications for human health are profound. Scientists hope to identify dysfunctions in specific neural circuits that could lead to diseases like Alzheimer's, Parkinson's or autism spectrum disorders. By better understanding these conditions at a microscopic level, treatments could be developed that target the root causes of neurological disorders.
Reimer and collaborators explain the different ways that someone can use the annotated 3D graphs produced by NEURD. For example, to categorize different classes of neurons and then calculate the probabilities of forming connections between these classes.
"We could ask question about connectivity between classes or about different spine densities. All these features are precomputed, so you just have to think of the questions and the data is all there to be found," he said. "The scale of this research is incredibly massive, and even though it is from a mouse brain, being able to interpret this data in this way is going to give us an entirely new view of the neuronal connections in our brains."
While more work is needed and researchers still are far from mapping the human brain in full detail, the progress made is transformative.
"We're entering a new era in neuroscience, where machine learning can be integrated with biological understanding to unlock the brain's mysteries. Our hope is that NEURD will help get more researchers working to extract insights from these massive data sets" Reimer said.
This study is a collaboration between researchers, for a full list of names, institutions, contributions and fundings click here.
