Protein abundance is defined by transcriptional, post-transcriptional and post-translational regulatory mechanisms. Understanding the code for gene expression could inform novel therapies. Here, we developed a machine learning pipeline, termed SONAR, to decipher the endogenous sequence code that defines the abundance of protein in human cells. SONAR predicts up to 63% of protein abundance independently of promoter or enhancer information. Our analysis reveals a strong - yet dynamic - cell-type specific sequence code. The deep knowledge of SONAR provides a map of biologically active sequence features (SFs), which we leveraged to manipulate protein expression and tailored to a specific cell-type. Beyond providing fundamental insights in gene expression regulation, our study offers novel means to improve therapeutic and biotechnology applications. Datasets of the protein models (with gamma=1) are included in this repository. Please refer to GitHub for more details on how the models were generated.