Using Deep Learning to Extrapolate Protein Expression Measurements.

Area of Science:

• Proteomics
• Computational Biology
• Bioinformatics

Background:

• Mass spectrometry (MS)-based proteomics typically quantifies only a fraction of the proteome.
• Existing computational methods for missing value imputation in proteomics have limitations.
• A need exists for in silico methods to estimate protein abundances across all proteins.

Purpose of the Study:

• To develop a novel deep learning method for comprehensive protein abundance prediction in label-free MS experiments.
• To leverage gene functional annotations and RNA measurements as predictive features.
• To assess the performance and transferability of the developed method across different datasets and species.

Main Methods:

• A deep learning model was developed to extrapolate observed protein expression values.
• The model utilizes gene functional annotations and RNA expression data.
• The method was validated on four diverse datasets (human cell lines, human and mouse tissues).

Main Results:

• The proposed deep learning method achieved average R-squared scores between 0.46 and 0.54, outperforming RNA-based correlation methods.
• Models demonstrated successful transferability across experiments and species, with a human tissue model performing well on mouse tissue data.
• The predicted protein abundances can identify aberrant protein expression levels.

Conclusions:

• Computational prediction of protein abundances in label-free MS is feasible using functional annotations and RNA data.
• The developed deep learning approach offers a powerful tool for comprehensive proteome analysis.
• This method has the potential to enhance the interpretation of proteomic data and identify biological variations.