Journal DOI 10.68956/journal.human-performance-futures

Creative Commons source reading · Prediction and forecasting

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

McGuire CE; Hibbs MA

Originally published in PLoS One · 2026-04-16

From the article

Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.

Abstract

AbstractAutomated 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…

4. Discussion

In this work, we set out to compare feedforward neural networks to two classical machine learning methodologies for co-expression-based automated function prediction in yeast. We created a model of AFP in yeast using neural networks that can predict the functional relationship between pairs of genes in the context of 79 biological processes. Our model uses these pairwise predictions to rank individual genes by their predicted involvement for these 79 different biological processes. When training our model with the same gene expression input data and Gene Ontology labels as SPELL and MEFIT, we found that our model’s predictions of genes involved in mitochondrion organization had better ROC performance than these models. This indicates that if our model had been created at the same time as MEFIT and SPELL, it could have been used in…

5. Future work and conclusions

Our future work consists of improving the interpretability of our model, exploring other deep learning architectures, and integrating data modalities other than gene expression into our methodology comparisons. First, while our model displayed better predictive accuracy than MEFIT and SPELL for mitochondrion organization, we note that our neural network model is less explainable and interpretable than these other methods. Given a predicted gene function, MEFIT and SPELL both display the input gene expression datasets that were most important for making that prediction [4,5]. Our model lacks this explainability. We did improve the interpretability of our model by finding that a GO term’s performance is correlated with the extent the co-expression of a term’s gene pairs diverge from the background training distribution (S4 Fig in S1…

Editorial reading guide

  • What data created and tested the model, and was evaluation independent?
  • What is the cost of a false positive, false negative or out-of-distribution decision?