Joint learning improves protein abundance prediction in cancers

Area of Science:

• Molecular Biology
• Genomics
• Proteomics

Background:

• The central dogma (DNA to mRNA to protein) is complicated by regulatory mechanisms affecting protein translation.
• Weak correlations between mRNA and protein levels are common in cancer and across samples.
• Accurate prediction of protein abundance from RNA is crucial for understanding cancer biology.

Purpose of the Study:

• To develop a robust method for predicting proteome from transcriptome data.
• To improve the accuracy of protein abundance prediction by integrating multi-cancer data.
• To identify key regulatory pathways and network modules influencing proteomic profiles in cancer.

Main Methods:

• Developed a generic model for single-gene mRNA-protein correlation.
• Built gene-specific models capturing interdependencies within regulatory networks.
• Created a cross-tissue model by jointly learning shared pathways across breast and ovarian cancer samples.
• Utilized NCI-CPTAC and TCGA datasets for training and validation.

Main Results:

• The developed method ranked first in the NCI-CPTAC DREAM Proteogenomics Challenge.
• Predictive performance approached the accuracy of experimental replicates.
• Identified key functional pathways and network modules controlling proteomic abundance in cancers.
• Highlighted the importance of metabolism-related genes in cancer proteomic regulation.

Conclusions:

• Presented a novel method to predict proteome from transcriptome using a trans-tissue model.
• Demonstrated the value of integrating multi-cancer information for improved prediction.
• Provided a foundation for future research in cancer proteogenomics.