Proteins are made when ribosomes move along messenger RNA (mRNA) molecules, reading their genetic instructions one codon at a time. Exactly where ribosomes pause or move quickly can influence how much protein is produced, making ribosome behavior an important part of gene regulation. Traditionally, researchers have relied on specialized experiments such as ribosome profiling, RNA sequencing, and genomic information to study this process.

Researchers from Carnegie Mellon University have developed a new computational framework called seq2ribo that predicts ribosome positions using only an RNA sequence as input.

Overview of seq2ribo during inference

Given an input RNA sequence, we first derive codon-level structural features and pass the sequence and features into the sTASEP module of seq2ribo, which simulates a raw ribosome location profile. The polisher module of seq2ribo then combines the codon sequence, structural features, and simulated profile to produce a polished ribosome location profile. We reuse this shared polished profile for two sequence-only downstream tasks. We predict translation efficiency (TE) for human transcripts and predict protein expression, and we can also use the predicted ribosome profiles to generate synthetic ribosome profiling data. The bottom panels summarize how we fit the sTASEP parameters, train the polisher model, and finetune seq2ribo for the downstream regression tasks.

Current approaches often depend on experimental datasets or simplified computer simulations that cannot fully capture the complexity of protein translation. This makes it difficult to design entirely new RNA sequences for applications such as mRNA vaccines or synthetic biology.

Seq2ribo addresses this challenge by combining a biological simulation with artificial intelligence. The framework first uses a structure-aware simulation that models how ribosomes move along an RNA molecule while accounting for factors such as codon usage, RNA structure, local base pairing, and other sequence characteristics. A machine learning model then refines these predictions to more closely match experimentally observed ribosome behavior.

The researchers evaluated seq2ribo using multiple human cell types, including induced pluripotent stem cells, HEK293 cells, lymphoblastoid cell lines, and retinal pigment epithelial cells. Across these datasets, the framework consistently outperformed existing sequence-based methods in predicting where ribosomes are located along RNA molecules.

In addition to predicting ribosome positions, the researchers showed that seq2ribo can estimate translation efficiency and protein expression with high accuracy. This means the software not only predicts where ribosomes bind, but also provides insight into how efficiently an RNA molecule may produce protein.

One of the biggest advantages of seq2ribo is that it works directly from RNA sequence alone. Because it does not require experimental expression data or complete genomic context, it could become a valuable tool for designing new RNA molecules before they are synthesized and tested in the laboratory.

As RNA sequencing continues to expand our understanding of gene expression, computational tools such as seq2ribo will help bridge the gap between RNA sequence and protein production. Predicting ribosome behavior directly from sequence could accelerate the development of optimized mRNA therapeutics, vaccines, and other synthetic biology applications.

Availability: seq2ribo is available at https://github.com/Kingsford-Group/seq2ribo.