Evolutionarily informed deep learning methods for predicting relative transcript abundance from DNA sequence.

Area of Science:

• Molecular Biology
• Genetics
• Bioinformatics
• Machine Learning

Background:

• Deep learning shows promise in molecular biology and genetics.
• Current deep learning models overlook evolutionary dependencies, leading to potential inaccuracies.
• Integrating evolutionary information is crucial for reliable biological predictions.

Purpose of the Study:

• To develop and validate novel machine learning approaches that account for evolutionary relatedness in biological data.
• To improve the accuracy and interpretability of predictive models in molecular biology.
• To investigate the role of evolutionary context in mRNA expression level prediction.

Main Methods:

• Developed gene-family-guided splitting to ensure distinct gene families in training and testing sets.
• Implemented ortholog contrasts for evolutionarily informed comparisons between homologous genes.
• Validated approaches using mRNA expression level prediction tasks.

Main Results:

• Achieved area under the ROC curve (auROC) values between 0.75 and 0.94 in mRNA expression prediction.
• Model weight inspections revealed biologically interpretable patterns.
• Identified distinct roles for 5' and 3' untranslated regions (UTRs) in regulating mRNA abundance.

Conclusions:

• Machine learning models incorporating evolutionary relatedness significantly improve prediction accuracy in molecular biology.
• Gene-family-guided splitting and ortholog contrasts are effective strategies for integrating evolutionary information.
• The 3' UTR plays a key role in fine-tuning mRNA levels, while the 5' UTR influences large-scale changes, offering new hypotheses for gene regulation.