Revisiting co-expression-based automated function prediction in yeast with neural networks and updated Gene Ontology annotations

Automated function prediction (AFP) is the process of predicting the function of genes or proteins with machine learning models trained on high-throughput biological data. Deep learning with neural networks has become the dominant machine learning methodology of contemporary AFP models. However, it is unclear what difference exists between neural networks and classical machine learning techniques for AFP. Therefore, we created a model of AFP in yeast using feedforward neural networks that is trained on gene co-expression data to predict Gene Ontology (GO) labels and directly compared it to two previous, experimentally-validated AFP models. When trained on the same input data, our model outperforms these classical machine learning models (a Bayesian network and adaptive query-driven search) when predicting individual genes involved in mitochondrion organization. In particular, we found our neural network model better distinguished mis-annotated negatives in its training data. Finally, we quantified how differences in gene expression data and Gene Ontology annotations affect the performance of our model across each of its predicted GO terms. Our results suggest that feedforward neural networks can be more performant and robust to GO mis-annotations compared to the specific classical approaches examined here for co-expression-based AFP of some biological processes.