The increasing demand for biological products drives many efforts to engineer cells that produce heterologous proteins at maximal yield. Recent advances in massively parallel reporter assays can deliver data suitable for training machine learning models and sup-port the design of microbial strains with optimized protein expression phenotypes. The best performing sequence- to-expression models have been trained on one-hot encodings, a mechanism-agnostic representation of nucleotide sequences. Despite their excellent local pre-dictive power, however, such models suffer from a limited ability to generalize predictions far away from the training data. Here, we show that libraries of genetic constructs can have substantially different cluster structure depending on the chosen sequence representation, and demonstrate that such differences can be leveraged to improve generalization perfor-mance. Using a large sequence- to-expression dataset fromEscherichia coli, we show that non-deep regressors and convolutional neural networks trained on one-hot encodings fail to generalize predictions, and that learned representations using state-of-the-art large language models also struggle with out-of-domain accuracy. In contrast, we show that despite their poorer local performance, mechanistic sequence features such as codon bias, nucleotide con-tent or mRNA stability, provide promising gains on model generalization. We explore several strategies to integrate different feature sets into a single predictive model, including feature stacking, ensemble model stacking, and geometric stacking, a novel architecture based on graph convolutional neural networks. Our work suggests that integration of domain-agnostic and domain-aware sequence features offers an unexplored route for improving the quality of sequence- to-expression models and facilitate their adoption in the biotechnology and phar-maceutical sectors.