seq2ribo: Structure-aware integration of machine learning and simulation to predict ribosome location profiles from RNA sequences

The paper introduces seq2ribo, a hybrid framework combining a structure-aware TASEP simulation with a machine learning polisher to accurately predict ribosome location profiles and protein expression directly from mRNA sequences, thereby enabling de novo mRNA design without reliance on experimental data or genomic context.

The Gist

Imagine your cell is a bustling factory. The DNA is the master blueprint stored in the office, but the mRNA is the photocopied work order sent out to the factory floor. The ribosomes are the workers on the assembly line, reading the work order and building proteins (the final product).

Sometimes, the assembly line moves smoothly. Other times, it gets clogged. Workers might pause to tie their shoes, get stuck in a narrow hallway, or wait for a specific tool. These "traffic jams" determine how many products get made and how well they are folded.

For a long time, scientists could only see these traffic jams by actually stopping the factory and taking a snapshot (a technique called Ribo-seq). But what if you wanted to design a new work order for a vaccine or a medicine before you even built it? You couldn't take a snapshot of something that doesn't exist yet. You needed a way to predict the traffic jams just by looking at the text of the work order.

Enter seq2ribo, a new tool created by researchers at Carnegie Mellon University. Think of it as a super-smart traffic simulator that can predict exactly where the workers will get stuck, just by reading the mRNA sequence.

Here is how it works, broken down into simple steps:

1. The "Rough Draft" Simulator (sTASEP)

Imagine you are trying to predict traffic on a highway. You know that some cars are slow (like heavy trucks) and some roads are narrow.

• The Old Way: Scientists used a simple model that only looked at the "cars" (the genetic code). They assumed every truck took the same amount of time to pass a toll booth.
• The seq2ribo Way: The researchers built a smarter simulator called sTASEP. This simulator knows that the highway isn't just a straight line; it has folds and loops (like a paper crane).
  • It knows that if the road folds into a tight knot, the workers (ribosomes) will get stuck.
  • It knows that some workers are faster than others.
  • It runs a simulation to create a "rough draft" of where the traffic jams will happen.

2. The "Editor" (The Polisher)

Even the best rough draft has mistakes. The simulator might get the general idea right (a jam here, a smooth road there), but it might not be 100% accurate.

• This is where the Polisher comes in. Think of this as a highly experienced editor who uses Artificial Intelligence (specifically a type called Mamba).
• The editor looks at the "rough draft" from the simulator, reads the original text again, and says, "Ah, I see the simulator thought the workers would pause here, but based on the shape of the road, they'll actually pause two spots to the left."
• It refines the prediction, making it incredibly precise.

Why is this a Big Deal?

1. It works on "Invisible" Sequences
Previously, to know how well a protein would be made, you had to actually make it in a lab and measure it. With seq2ribo, you can type in a brand-new sequence (like a new mRNA vaccine design) and instantly see a map of where the ribosomes will go. It's like having a weather forecast for a storm that hasn't happened yet.

2. It's a "Two-Step" Dance
The magic is in combining physics (the simulator) with learning (the AI).

• The simulator provides the "common sense" rules of how traffic works.
• The AI learns the subtle, weird exceptions that the rules miss.
• Together, they are much better than either could be alone.

3. It Predicts the "Output"
The researchers didn't just stop at predicting traffic jams. They showed that if you know where the traffic jams are, you can predict:

• Translation Efficiency: How much product will be made? (The answer: If there are fewer jams, you get more product).
• Protein Expression: How much of the final protein will actually appear in the cell?
• In their tests, seq2ribo was able to predict protein production with 90% accuracy, beating all previous methods.

The Real-World Impact

Imagine you are designing a new mRNA vaccine.

• Before: You design a sequence, synthesize it, test it in a lab, find out it's too slow or makes too little protein, and then try again. This takes months.
• With seq2ribo: You design 1,000 different sequences on a computer. The tool simulates the traffic for all of them in minutes. You pick the top 10 that look like they will have the smoothest assembly lines, synthesize only those, and test them.

In short: seq2ribo is a crystal ball for molecular biology. It turns the complex, chaotic dance of ribosomes into a predictable map, allowing scientists to design better medicines and vaccines faster and cheaper than ever before.

Technical Summary

1. Problem Statement

The translation of mRNA into protein is a dynamic process governed by initiation, elongation, and termination, heavily influenced by ribosome pausing, traffic jams, and mRNA secondary structures. Accurately predicting ribosome occupancy profiles (Ribo-seq data) is crucial for understanding protein expression levels and designing synthetic mRNA therapeutics (e.g., vaccines).

Current approaches face significant limitations:

• Experimental Dependency: Most tools rely on existing Ribo-seq or RNA-seq data, making them unsuitable for de novo sequence design where no experimental data exists.
• Oversimplified Simulation: Mechanistic models like the Totally Asymmetric Simple Exclusion Process (TASEP) often ignore mRNA secondary structure, focusing only on codon elongation times, which leads to poor fidelity in predicting positional ribosome distributions.
• Pure Machine Learning Limitations: Deep learning models often require experimental labels for training or fail to capture the physical dynamics of ribosome traffic without explicit mechanistic priors.

The core challenge is to develop a method that predicts high-fidelity, codon-level ribosome location profiles using only the mRNA sequence as input, without requiring genomic context or experimental expression data.

2. Methodology

The authors propose seq2ribo, a hybrid framework combining mechanistic simulation with deep learning. The pipeline consists of two main stages:

A. Structure-Aware TASEP (sTASEP) Simulator

The first stage is a stochastic simulator that generates a "raw" ribosome profile. Unlike classical TASEP, sTASEP incorporates structural and positional features into the elongation rate calculation.

• Mechanism: Ribosomes move on a 1D lattice of codons. Movement is blocked if the exclusion zone (3 codons upstream/downstream) is occupied.
• Elongation Rate ($r_j$): The rate at a codon $j$ is determined by a wait time $w_j$, calculated as:
  $$w_j = \tau(C_j) + \alpha p_j + \beta a_j + \gamma b_j$$
  Where:
  • $\tau(C_j)$: Codon-specific wait time (61 parameters for non-stop codons).
  • $p_j$: Pair wait time based on the number of nucleotides in the codon involved in base pairing (0–3).
  • $a_j$: Angle wait time based on discretized changes in the local backbone angle (4 bins).
  • $b_j$: Bucket wait time based on the relative position of the codon in the CDS (first, middle, or last third; 3 bins).
• Training: Parameters are fitted to minimize the scaled Mean Absolute Error (MAE) between simulated and observed profiles, ensuring the model captures positional distribution rather than just total abundance.

B. Mamba-Based Polisher

The second stage refines the raw simulation output using a deep learning model based on Mamba (a Structured State Space Model).

• Input: The polisher receives five channels: the codon sequence, the pair sequence, the angle sequence, the bucket sequence, and the simulated A-site counts from sTASEP.
• Architecture: It embeds these inputs into a latent space and processes them through stacked Mamba blocks (Layer Norm + Mamba layer + Residual connection).
• Output: A "polished" codon-level ribosome count profile.
• Loss Function: Trained using Poisson negative log-likelihood to match observed A-site counts.

C. Downstream Tasks

The framework includes task-specific heads for:

1. Translation Efficiency (TE) Prediction: Aggregating the polished profile to predict TE.
2. Protein Expression Prediction: Predicting protein output levels.
3. Synthetic Ribo-seq Generation: Generating in silico profiling data for arbitrary sequences.

3. Key Contributions

1. First Sequence-Only High-Fidelity Predictor: seq2ribo is the first method to achieve meaningful positional correlation with observed ribosome profiles using only the RNA sequence, outperforming all baselines that rely on genomic context or experimental data.
2. Hybrid Architecture: It successfully bridges the gap between mechanistic physics (sTASEP) and data-driven learning (Mamba), leveraging the interpretability of simulation and the flexibility of deep learning.
3. Structure-Aware Simulation: The introduction of sTASEP, which explicitly models local base pairing, structural angles, and positional buckets, significantly improves upon classical TASEP.
4. Generalizability: The method was validated across four distinct cell lines (iPSC, HEK293, LCL, RPE-1) and demonstrated strong transferability to protein expression prediction tasks.

4. Results

The model was evaluated on four cell lines using nine metrics, including Pearson correlations (Tx-level $r$, Shape $r$, Elemwise $r$) and various MAE metrics.

• Ribosome Profile Prediction:

  • Correlation: seq2ribo achieved transcript-level Pearson correlations ($r$) between 0.657 and 0.920, and within-transcript shape correlations up to 0.186.
  • Baseline Comparison: All baselines (classical TASEP, sequence-only Translatomer) yielded near-zero or negative shape correlations. seq2ribo is the only method to capture the positional distribution of ribosomes.
  • Error Reduction: It reduced element-wise MAE by 30.3% to 37.7% relative to the Translatomer baseline.
  • Structural Metrics: sTASEP alone reduced codon-level MAE by up to 95.6% compared to classical TASEP.

• Downstream Task Performance:

  • Translation Efficiency (TE): In CDS-only settings, seq2ribo outperformed the RiboNN baseline across all tasks, achieving Pearson correlations up to 0.732 (with UTRs).
  • Protein Expression: After fine-tuning, seq2ribo achieved Pearson correlations between 0.830 and 0.903 on an external mRFP protein expression dataset, surpassing the performance of CodonBERT.

• Ablation Studies:

  • Removing the sTASEP simulation reduced performance, confirming that the mechanistic prior provides complementary information to the sequence and structural features.
  • The full hybrid model (Sequence + Structure + sTASEP) consistently outperformed any single component.

5. Significance

• Synthetic Biology & mRNA Design: By operating solely on sequence, seq2ribo enables the rational design and optimization of mRNA therapeutics (e.g., vaccines) without the need for prior experimental expression data or specific genomic context.
• Synthetic Data Generation: It serves as a generator for synthetic Ribo-seq data, allowing researchers to screen candidate mRNA libraries in silico to optimize ribosome traffic patterns (e.g., minimizing collisions) before physical synthesis.
• Biological Insight: The fitted parameters of sTASEP are biologically interpretable, showing correlations with Codon Stabilization Coefficients (CSC) and consistent wait times across cell lines, validating the model's biological relevance.
• Methodological Advancement: The work demonstrates that coupling mechanistic simulation with modern state-space models (Mamba) is a powerful paradigm for modeling complex biological dynamics where pure data-driven approaches struggle due to data scarcity or lack of physical constraints.

Availability: The code and pretrained models are available at https://github.com/Kingsford-Group/seq2ribo.