Deep Learning Facilitates Understanding of Genetic Interactions

2019-12-11

A team of computer scientists from Carnegie Mellon University has used a deep learning approach that has revolutionized facial recognition and other image-based applications in recent years, redirecting its capabilities towards exploring relationships between genes.


They say the trick lies in transforming a large amount of gene expression data into images, which can then be analyzed by convolutional neural networks (CNNs) that are adept at analyzing visual images. CNNs outperform existing methods in this task.Researchers published a paper today in the Proceedings of the National Academy of Sciences on how CNNs help identify genes related to diseases and developmental and genetic pathways that could be potential drug targets. However, Ziv Bar-Joseph, a professor of computational biology and machine learning, says the application of the new method called CNNC may go far beyond gene interactions.


Bar-Joseph said, "CNNs, which were developed ten years ago, have had a revolutionary impact. Sometimes we take this technology for granted because we have been using it all along. But it is powerful and not limited to images. It all depends on how you represent the data."


In this case, Bar-Joseph and other researchers are studying gene relationships. Humans have approximately 20,000 genes that work together, so it is necessary to understand how genes collaborate in complexes or networks to understand questions related to human development or diseases.One way to infer these relationships is by looking at gene expression, which represents the activity levels of genes in cells. Typically, if gene A is activated, gene B is also activated, which is a clue to their interaction. However, this could be a coincidence or both could be activated by a third gene C. Several previous methods have been developed to determine these relationships.


To use CNNs to help analyze gene relationships, Bar-Joseph used single-cell expression data, which experiments can determine the levels of each gene in individual cells. Then, the results of thousands of these single-cell analyses were arranged in the form of a matrix or histogram, so that each cell of the matrix represents different co-expression levels of a pair of genes.


Presenting the data in this way adds a spatial aspect, making the data more like images and therefore easier for CNNs to use. By using gene data that already have established interactions, researchers were able to train CNNs to recognize which genes interact and which do not based on visual patterns in the data matrix.Yuan, a researcher, said, "Distinguishing causality from correlation is very, very difficult, but the CNNC method is statistically more accurate than existing methods. He and Bar-Joseph expect CNNC to become one of several techniques that researchers will ultimately use to analyze large datasets."


Bar-Joseph said, "This is a very general method that can be applied to many analyses. The main limitation is the data—the more data, the better CNNs work. Cell biology is very suitable for using CNNC because typical experiments may involve thousands of cells and generate a large amount of data."

RECENT AI NEWS

RECENT AI TOOLS