Deep learning identifies mosaic mutations that cause illness.
A deep learning method developed by researchers at the University of California, San Diego (UCSD) School of Medicine and Rady Children's Genomic Medicine Institute may detect disease-causing mosaic mutations. According to the experts, this is the first step in developing treatments for a variety of ailments.
Mosaic mutations are found in just a small fraction of human cells, making them challenging to identify using traditional DNA sequencing and computational tools.
To address this issue, researchers used deep learning to create a technology that can learn from massive volumes of genetic data and sift through it to identify the small fraction of cells impacted by mutations. The potency of this "DeepMosaic" technique is outlined in a paper published yesterday in the journal Nature Biotechnology.
DeepMosaic was trained on 180,000 simulated or experimentally tested MVs and compared to 619,740 simulated MVs and 530 independent MVs produced from biological studies of 16 genomes and 181 exomes. The tool functions similarly to human visual processing, but with considerably more precision and attention to detail, as well as enhanced computational approaches for detecting non-cancerous MVs, which are frequently missed.
"Focal epilepsy is an example of an undiagnosed condition," says Joseph Gleeson, a neuroscience professor at the University of California, San Diego School of Medicine, the head of neuroscience research at the Rady Children's Institute for Genomic Medicine, and the paper's primary author.
"Epilepsy affects four percent of the population, and around a quarter of focal seizures do not respond to traditional treatments. To halt the seizures in these individuals, the focal portion of the brain is frequently removed. Mosaic mutations in the brain can produce focal seizures in some people.
"We had numerous epileptic cases for whom we couldn't pinpoint a reason, but when we used our DeepMosaic approach to the genetic data, the mutation became clear. This enabled us to improve the sensitivity of DNA sequencing in some types of epilepsy, leading to findings that offer the possibility of novel therapies for brain illnesses."
DeepMosaic's training data included known SNs as well as many regular DNA sequences, allowing the machine learning approach to understand the distinctions between these sequences. The tool's development needed an iterative process of constant retraining with more complicated information and the selection of hundreds of models.
The computer outperformed the human eye and prior approaches in detecting mosaic mutations. It was also tested on huge independent datasets that had not been used for training and outperformed earlier analysis and detection approaches.
DeepMosaic had a sensitivity of 0.78, specificity of 0.83, and positive predictive value of 0.96 for noncancer whole-genome sequencing data in their research, and it twice the validation rate compared to the best prior techniques for noncancer whole-genome sequencing data (0.43 versus 0.18).
"DeepMosaic outperformed standard techniques in identifying mosaicism from genomic and exon sequences," says Xin Xu, co-author of the original article and former assistant professor at UC San Diego School of Medicine. "The deep learning models capture distinguishing visual cues that are quite comparable to those that experts focus on during manual variant screening."
DeepMosaic, the researchers concluded, is a reliable TM classifier for non-cancerous samples that may be used as an alternative or addition to existing approaches.
To encourage additional researchers to adopt DeepMosaic, the UCSD team and the Council made it publicly available through an open-source platform that will allow other researchers to train neural networks for more focused mutation identification using a similar image-based technique.
Spatial transcriptomics and proteomics are combined with a unique high-throughput genomics technique.
Researchers have greatly increased the precision of differential gene expression analyses in various tissue locations by merging spatial transcriptomics with protein markers. SPOTS records the activity of distinct cell types in far higher detail than unimodal measures. This approach generates data-rich organ and tumor maps.
The study included Weill Cornell Medicine, NewYork-Presbyterian, and the New York Genome Center researchers. The study was published in the journal Nature Biotechnology this week.
"This technique is exciting because it allows us to map the spatial organization of tissues, including cell types, cellular activities, and cell-cell interactions, in unprecedented detail," says study co-author Dan Landau, an associate professor in Weill Cornell Medicine's Department of Hematology and Medical Oncology, a member of the Sandra and Edward Meyer Cancer Center, and a tenured faculty member at the New York Genome Center.
One of the most rapidly increasing areas of study is spatial transcriptomics, which shows gene expression and location. Current approaches, however, frequently include dissolving tissues and isolating cells from their neighbors, resulting in a loss of information regarding the initial position of profiled cells in the tissue. The SPOTS technology, according to the researchers, gathers this geographical information with excellent resolution.
The SPOTS system is built in part on 10x Genomics' current technology. It employs glass slides that may be used to image tissue samples using classic microscopic pathology procedures. The slides, on the other hand, are coated with thousands of unique probes, each of which has a chemical barcode indicating the two-dimensional location on the item.
When a thin portion of a tissue sample is put on the slide and the cells become permeable, the probe molecules pick up messenger RNA, or transcripts of activated genes, from surrounding cells.
The modified antibodies bind to tissue proteins as well as the probe molecules of interest. Researchers may then identify the acquired mRNA and target proteins and put them exactly where they belong in the tissue sample. The maps produced may then be compared to a conventional pathological picture of the sample.
The researchers used SPOTS to examine splenic tissue from normal mice and uncovered clusters of distinct cell types, their functional states, and how these states fluctuate depending on cell location.
It was also employed to map the cellular arrangement of a maternal mouse tumor by the researchers. The map revealed that macrophages were in two states, as indicated by protein markers: one active and fighting the tumor, and the other immunosuppressive and forming a barrier to protect the tumor.
"We found that these two subgroups of macrophages are localized in distinct parts of the tumor and interact with different cells, and this diversity in microenvironment probably explains their varied levels of activity," Landau, an oncologist, adds.
Because the first version of SPOTS has such great spatial resolution, each "pixel" in the generated dataset summarizes information about at least a few cells' genetic activity. According to Landau, the researchers aim to soon be able to lower the resolution to individual cells while adding extra layers holding critical biological information.