Neural Networks for Gene Function Prediction in Yeast Mitochondrial Organization

The increased accessibility of high-throughput biological experimentation has promised to revolutionize our understanding of cellular systems on a large scale. Bioinformatics enables the use of computer algorithms and machine learning to aid in our ability to interpret these biological data and increase the rate at which we make novel biological discoveries. Over the past two decades, there have been many attempts to use machine learning to predict gene function in Saccharomyces cerevisiae, or baker's yeast, using mitochondrial organization as a model system. While these past methods were successful, newer machine learning methods may demonstrate even better results. The goal of this study was to assess the viability of neural networks as a tool for gene function prediction and determine what unique advantages they may provide in comparison to other machine learning methods. Therefore, we created a model of gene function prediction in yeast that uses a neural network to predict a gene's involvement in 51 unique biological processes. Our model uses publicly available gene expression data obtained from Saccharomyces Genome Database to predict the functional relationship between pairs of genes in the context of 51 Gene Ontology terms. The model makes predictions by training a fully-connected linear neural network on labeled training data of gene pairs known to be functionally related. We then compared our model's performance to that of the models bioPIXIE, MEFIT, and SPELL. We trained our network using the same input data used by these models, then compared all four models' predictions of genes involved in mitochondrion organization. We found that neural networks perform better than bioPIXIE, MEFIT, and SPELL when training on the same input data. Furthermore, we found that our model's performance can be partially attributed to the ability of a neural network to exploit shared information between biological processes when predicting multiple Gene Ontology terms simultaneously. In conclusion, our results suggest that neural networks are a more accurate model of gene function prediction than previous machine learning models. Furthermore, gene function prediction with neural networks has promising potential for accelerating the rate at which researchers can interpret biological data and make novel biological discoveries. This research was funded by the HEP Computer Science Fellowship.