Electronic device and operating method for generating gene feature vectors for distributed gene representation based on correlations between genes associated with cancer

Apparatus for Generating Gene Feature Vectors Based on Prognosis-Specific Gene Correlation Network Learning

This technology is an electronic device that generates feature vectors for distributed gene representation by calibrating gene networks using genetic data from cancer patients with good and poor prognoses and training them at the pathway level. It represents each gene as a numerical vector by training a weight matrix on pathway vectors from a calibrated network where low-correlation links have been removed.

Predicting cancer prognosis based on genetic characteristics requires representing abstract gene features as numerical values that computers can understand, but methods for this have not been sufficiently established. Even when applying distributed representation concepts from natural language processing to genes, there have been limitations in reflecting the interaction relationships between genes according to prognosis.

This technology calculates the correlation coefficients of gene pairs for each prognosis group to filter out links with weak actual interactions and uses continuous link paths as training data to obtain gene feature vectors that incorporate prognostic information. By removing redundant paths common to both groups, features that distinguish prognosis become more distinct. It can be applied to genome-based cancer prognosis prediction models and bioinformatics analysis tools, fundamentally improving the quality of input features for prediction models.

‍

‍

Key Features:

• A correlation coefficient calculation unit that computes first and second correlation coefficients for each gene pair connected in a gene network using gene data from good and poor prognosis groups
• A calibrated gene network generation unit that creates first and second calibrated gene networks by removing links between gene pairs whose correlation coefficients are below a reference value
• A path vector generation unit that generates path information for genes connected by n consecutive links in the calibrated gene network and converts them into path vectors
• A gene feature vector determination unit that determines gene feature vectors using a weight matrix trained so that the output value of the path vector subjected to weight matrix calculation approaches the result value for each prognosis group

‍

This invention was developed with support from the Ministry of Education for research on drug repurposing prediction techniques using network-based machine learning.

‍