Inferring protein from transcript abundances using convolutional neural networks.

Background: Although transcript abundance is often used as a proxy for protein abundance, it is an unreliable predictor. As proteins execute biological functions and their expression levels influence phenotypic outcomes, we developed a convolutional neural network (CNN) to predict protein abundances from mRNA abundances, protein sequence, and mRNA sequence in Homo sapiens (H. sapiens) and the reference plant Arabidopsis thaliana (A. thaliana).

Results: After hyperparameter optimization and initial data exploration, we implemented distinct training modules for value-based and sequence-based data. By analyzing the learned weights, we revealed common and organism-specific sequence features that influence protein-to-mRNA ratios (PTRs), including known and putative sequence motifs. Adding condition-specific protein interaction information identified genes correlated with many PTRs but did not improve predictions, likely due to insufficient data. The integrated model predicted protein abundance on unseen genes with a coefficient of determination (r²) of 0.30 in H. sapiens and 0.32 in A. thaliana.

Conclusions: For H. sapiens, our model improves prediction performance by nearly 50% compared to previous sequence-based approaches, and for A. thaliana it represents the first model of its kind. The model's learned motifs recapitulate known regulatory elements, supporting its utility in systems-level and hypothesis-driven research approaches related to protein regulation.