Inferring protein from transcript abundances using convolutional neural networks

Progress in personalized medicine and crop optimization hinges on quantitative, systems-level insights into how genotype and environment shape cellular phenotypes. Because proteins and their interactions underlie nearly all cellular functions, fluctuations in their concentrations strongly influence phenotypic outcomes. Yet protein measurements remain far more expensive than high-throughput mRNA assays, which are therefore often used as surrogates even though transcript and protein levels correlate only imperfectly. Recent advances in artificial intelligence offer a promising approach to reduce this inaccuracy. I therefore developed species-specific convolutional neural networks (CNNs) for Homo sapiens and Arabidopsis thaliana that predict protein abundances directly from paired transcript abundances and raw sequence data. Trained on matched transcriptome-proteome datasets, the models achieve coefficients of determination of 0.30 and 0.32, respectively. They improve accuracy for H. sapiens by about 40% over the best sequence-based approach and provide the first published model for A. thaliana. Analysis of the learned filter weights shows that the networks autonomously rediscover known regulatory motifs governing mRNA decay and translation while identifying several novel elements that merit experimental validation. An extended architecture that incorporates expression profiles of putative interaction partners does not yet surpass the sequence-only model, underscoring the need for larger, condition-rich training sets. I also developed an automated image-analysis pipeline for high-throughput yeast two-hybrid systems. Classical computer-vision algorithms locate 96-well plate grids, and a second CNN scores yeast colony growth, generating quantitative protein-protein interaction networks from thousands of colonies. Integrating these curated interaction data with graph-convolutional modules alongside sequence features processed by conventional convolutions offers a promising route to further improve the accuracy of protein abundance prediction. Collectively, these contributions demonstrate how CNNs can both extract and predict reliable biological information from raw experimental data, advancing systems-level modeling of gene expression and enabling scalable applications in personalized medicine and crop improvement.