Leveraging amino acid embeddings for accurate translation efficiency prediction

MOTIVATION: Accurate prediction of translation efficiency from mRNA sequences is an important step toward mRNA sequence design. In recent years, the widespread adoption of high-resolution ribosome profiling techniques, such as Ribo-seq, has led to rapid accumulation of large-scale datasets that capture the relationship between mRNA sequences and translation efficiency. This progress has facilitated the development of machine-learning-based models for translation efficiency prediction. However, most existing approaches rely primarily on mRNA-derived features, such as codon usage frequency and 5' untranslated regions (5'UTRs), and do not directly utilize the amino-acid-level information. RESULTS: In this study, we show that incorporating distributed representations (embeddings) of amino acid sequences translated from mRNA as model features can improve translation efficiency prediction. By combining fixed amino acid embeddings derived from a pretrained protein language model with mRNA-derived features, the proposed model integrates complementary protein-level and nucleotide-level information for more accurate prediction of translation efficiency. AVAILABILITY AND IMPLEMENTATION: Codes are available at GitHub (https://github.com/gterai/AA_emb) and Zenodo (https://doi.org/10.5281/zenodo.23968476).