Multi-dataset Integration and Residual Connections Improve Proteome Prediction from Transcriptomes using Deep...

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Proteome and transcriptome data frequently show poor correlation.
• Predicting protein quantities from gene expression (transcriptomics) is challenging.
• Understanding transcript-protein relationships is crucial in biological research.

Purpose of the Study:

• To enhance the prediction accuracy of proteome quantities from transcriptomic data.
• To investigate the impact of deep learning architectures on this prediction task.
• To identify functionally important transcripts for protein prediction.

Main Methods:

• Utilized publicly available data from the Clinical Proteomics Tumor Analysis Consortium (CPTAC).
• Employed deep learning models developed through neural architecture search (NAS).
• Applied model interpretation techniques (SHAP) to analyze transcript importance.

Main Results:

• Deep learning models, particularly those with residual connections, significantly improved proteome prediction accuracy.
• Residual connections were found to be crucial for retaining input information within the network.
• SHAP analysis identified specific transcript groups vital for accurate protein level prediction.

Conclusions:

• Neural architecture search-driven deep learning offers a powerful approach for predicting proteomes from transcriptomes.
• Architectural choices, like residual connections, critically influence model performance.
• Model interpretability methods can reveal biological insights into gene-protein relationships.