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

Proteomes are well known to poorly correlate with transcriptomes measured from the same sample. While connected, the complex processes that impact the relationships between transcript and protein quantities remains an open research topic. Many studies have attempted to predict proteomes from transcriptomes with limited success. Here we use publicly available data from the Clinical Proteomics Tumor Analysis Consortium to show that deep learning models designed by neural architecture search (NAS) achieve improved prediction accuracy of proteome quantities from transcriptomics. We find that this benefit is largely due to including a residual connection in the architecture that allows input information to be remembered near the end of the network. Finally, we explore which groups of transcripts are functionally important for protein prediction using model interpretation with SHAP.