A Transformer-based Multi-omics Model for Translation Efficiency in S. cerevisiae

Precise regulation of protein synthesis is fundamental to cellular homeostasis and remains a primary target for synthetic biology applications. However, the non-linear relationship between mRNA abundance and protein levels presents complexities that poses challenges for predictive engineering. Here, we present TRIM, a Transformer-based RNA Inference Model that leverages full-length mRNA sequences and multi-omics data to predict translation efficiency. By employing a Parallel Expert Mixer, TRIM achieves robust prediction accuracy (R2 [≥] 0.8,Pearson r [≥] 0.9). Trained on multimodal data from massive Saccharomyces cerevisiae isolates, TRIM demonstrates outstanding biological interpretability, helping to decipher complex translational patterns such as synergistic effects between bases, sequence-dependent codon preference in different stages, and distinct attention on key secondary structures. These results indicate that the integration of multi-omics data with holistic sequence modeling can effectively decode the cis-regulatory grammar of translation as well as providing a scalable and interpretable generative framework for future synthetic biology engineering. Availability and ImplementationThe source code and data used to produce the results and analyses presented in the manuscript are available from Github (https://github.com/ZeusLiu666/TRIM).