Evaluation of machine learning models on protein level inference from prioritized RNA features

Area of Science:

• Genomics and Proteomics
• Bioinformatics and Computational Biology
• Biotechnology

Background:

• Transcriptome and proteome profiles often show discrepancies.
• Proteomic analysis is resource-intensive compared to transcriptomic analysis.
• Accurate prediction of protein levels from mRNA data is highly desirable.

Purpose of the Study:

• To evaluate machine learning models for inferring protein expression from RNA data.
• To benchmark predictive performance across diverse datasets and platforms.
• To identify optimal strategies for transcriptome-based protein level prediction.

Main Methods:

• Comprehensive evaluation of 13 machine learning models.
• Utilized 20 proteogenomic datasets (>2500 human samples, 13 tissues).
• Employed feature selection methods and ensemble modeling (voting ensemble).

Main Results:

• Machine learning models, especially voting ensembles, achieved excellent predictive performance.
• Feature selection combined with classical models enhanced prediction accuracy.
• Incorporating an mRNA proxy model further improved regression model performance.
• Dataset and gene characteristics influenced prediction outcomes.

Conclusions:

• Transcriptome-based protein level prediction is feasible and valuable.
• Voting ensemble models offer a robust approach for this task.
• This work provides practical insights for leveraging RNA data to infer protein profiles and understand biological functions.