HalluCodon enables species-specific codon optimization using multimodal language models

HalluCodon is a customizable framework that leverages fine-tuned multimodal language models and a hallucination-based design strategy to generate species-specific coding sequences that replicate natural codon usage patterns and enhance protein expression in diverse plant systems.

The Gist

Imagine you are a chef trying to cook a famous dish (a protein) in a brand-new kitchen (a plant cell). You have the perfect recipe (the amino acid sequence), but the kitchen staff (the plant's cellular machinery) speaks a slightly different dialect. If you write the recipe using the chef's original dialect, the staff might get confused, cook too slowly, or even burn the dish.

Codon optimization is the art of rewriting that recipe so the kitchen staff understands it perfectly, ensuring the dish is cooked quickly and deliciously.

This paper introduces HalluCodon, a new, super-smart AI tool designed specifically to rewrite these genetic recipes for plants. Here is how it works, broken down into simple concepts:

1. The Problem: The "Dialect" Barrier

In biology, DNA is like a language. Even though different plants (like corn, rice, or tobacco) all speak "DNA," they have different favorite words (codons) for the same ingredients.

• Old Way: Previous tools were like a dictionary that just picked the most common word for every ingredient. This is okay, but it ignores the context. It's like speaking in a monotone voice using only the most common words; it might be understood, but it sounds robotic and doesn't flow well.
• The Issue: Sometimes, using the "most common" word everywhere actually breaks the rhythm of the recipe, causing the plant to make a messy, misfolded protein.

2. The Solution: HalluCodon (The "Polyglot Chef")

HalluCodon is a customizable AI framework that acts like a master translator who doesn't just know the dictionary, but understands the culture and rhythm of the specific plant kitchen.

It uses two main "brains" (multimodal language models) to do the job:

• Brain 1: CodonNAT (The "Naturalness" Detective)
  • What it does: It reads the genetic recipe and asks, "Does this sound like a native plant sentence?"
  • The Analogy: Imagine a music critic listening to a song. CodonNAT checks if the rhythm and flow match what native plant genes sound like. It ensures the new recipe doesn't sound "foreign" or awkward to the plant's machinery.
• Brain 2: CodonEXP (The "Success" Predictor)
  • What it does: It predicts, "If we use this recipe, how much food (protein) will we actually get?"
  • The Analogy: This is like a business consultant looking at a plan and saying, "This strategy will make us a million dollars." It learns from real data to guess which genetic tweaks will lead to a bountiful harvest.

3. The Magic Trick: "Hallucination" Design

Most old tools used a "Genetic Algorithm," which is like a slow, trial-and-error process. Imagine trying to find the best route to a city by randomly driving around, checking a map, and hoping you get there. It takes forever.

HalluCodon uses Hallucination-based design.

• The Analogy: Instead of driving randomly, it's like having a GPS that can imagine the perfect route instantly. The AI "hallucinates" (generates) a new sequence based on what it knows works, then immediately checks if it's good.
• The Result: It finds the perfect genetic recipe 46 times faster than the old trial-and-error methods and produces better results.

4. The Secret Sauce: The "GC3" Balance

The researchers discovered a specific trick that plants love: using more letters G and C at the third spot of the genetic words (called GC3).

• The Analogy: Think of GC3 as adding extra stabilizers to a bridge. It makes the genetic message (mRNA) stronger and less likely to fall apart before it's used.
• The Catch: If you add too many stabilizers, the bridge becomes too heavy and hard to build (hard to synthesize in a lab) or gets blocked by security guards (methylation) that stop the plant from reading it.
• HalluCodon's Fix: It doesn't just max out the GC3. It finds the "Goldilocks zone"—adding just enough stability to make the protein flow, without making the recipe too heavy or triggering the plant's security systems.

5. The Proof: Cooking in the Lab

The team tested HalluCodon in tobacco plants (a common lab plant).

• They took a glowing red protein (DsRed2) and asked different tools to rewrite its recipe.
• The Result: The HalluCodon recipe made the plants glow 13 times brighter than the old standard method and significantly brighter than other high-tech AI tools.
• They also tested it on huge, complex proteins that usually fail to grow in plants. By using their "GC3 balancing" trick, HalluCodon successfully made these giant proteins appear, which other methods couldn't do.

Summary

HalluCodon is a next-generation tool that helps scientists design genetic recipes for plants. Instead of just swapping words randomly, it uses advanced AI to understand the plant's unique "language" and "rhythm." It creates recipes that are not only easy for the plant to read but also highly efficient, resulting in much higher yields of useful proteins for medicine, agriculture, and industry. It's like upgrading from a basic translator to a cultural expert who knows exactly how to make a plant happy and productive.

Technical Summary

1. Problem Statement

Codon optimization is critical for enhancing heterologous protein expression in plant synthetic biology, molecular farming, and transgenic crop development. However, existing strategies face several limitations:

• Oversimplification: Traditional methods (e.g., Codon Adaptation Index, CAI) rely solely on codon frequency statistics, ignoring codon context, mRNA secondary structure, and regulatory roles of rare codons.
• Data Scarcity & Generalization: Many deep learning models (e.g., CodonTransformer) are trained from scratch, requiring massive species-specific datasets. They struggle to generalize to new species without extensive retraining.
• Immune Response Differences: Unlike mammalian systems where codon optimization must avoid immune activation (e.g., in mRNA vaccines), plant optimization can be guided directly by species-specific genomic features without immune constraints, yet current tools do not fully leverage this.
• Computational Efficiency: Existing generative methods, particularly Genetic Algorithms (GAs), are computationally expensive and slow for iterative sequence design.

2. Methodology: The HalluCodon Framework

HalluCodon is a customizable framework that utilizes multimodal biological language models to design coding sequences (CDS) tailored to specific plant species. It consists of two core predictive modules and a generative strategy.

A. Core Modules

1. CodonNAT (Naturalness Evaluator):

   • Goal: Evaluates how well a CDS matches the host species' endogenous codon context.
   • Architecture: A joint fine-tuning of two pre-trained language models: ESM2 (protein language model, 650M parameters) and mRNA-FM (RNA language model).
   • Mechanism: Uses a Masked Language Modeling (MLM) approach. It learns the probability of specific codons occurring at specific amino acid positions based on the host's genomic context. It outputs a "Naturalness Score."
   • Training Data: Fine-tuned on the top 10% of highly expressed genes (based on Codon Similarity Index) from 15 plant species.

2. CodonEXP (Expression Predictor):

   • Goal: Predicts the probability that a CDS will result in high protein expression.
   • Architecture: Integrates features from CDS sequences, amino acid sequences (via ESM2), and experimental protein abundance data.
   • Mechanism: Formulated as a binary classification task (High vs. Low expression). It learns to link sequence features to protein abundance levels derived from the PaxDb database.
   • Performance: Achieved ~82–84% accuracy in predicting high expression in Maize, Rice, and Tobacco, outperforming models using only protein or nucleotide features.

B. Sequence Generation Strategy

The framework generates optimized sequences in two stages:

1. Initialization (CodonIni): CodonNAT predicts the most probable codon for each amino acid to create an initial CDS.
2. Iterative Optimization: The initial sequence is refined to maximize a Fitness Score ($F$), defined as the product of the Naturalness Score ($N$) and the Expression Probability Score ($E$):
   $$F = E \times N$$
   • CodonHa (Hallucination-based Design): The primary method. It uses gradient-guided iterative optimization. It calculates gradients of the expression probability with respect to codon embeddings and applies substitutions that improve expression while maintaining codon context (regularized by Jensen-Shannon Divergence against host usage).
   • CodonGa (Genetic Algorithm): A comparative method using crossover and mutation, which was found to be significantly slower and less efficient.

C. GC3 Optimization

Recognizing that high GC content at the third codon position (GC3) correlates with mRNA stability in plants, HalluCodon includes a "GC3-rewarding" variant (Ha-GC3) that selectively biases substitutions toward G or C at the third position, particularly beneficial for large proteins.

3. Key Contributions

• Multimodal Fine-Tuning: First framework to jointly fine-tune protein (ESM2) and RNA (mRNA-FM) language models for species-specific codon optimization, allowing transfer learning with limited data.
• Hallucination-Based Design: Demonstrated that gradient-guided "hallucination" design is superior to Genetic Algorithms for codon optimization, offering 46x faster computation and higher experimental yields.
• Species-Specific Customization: The framework is designed to be retrained by users with their own datasets, making it adaptable to any plant species with available omics data.
• Web Interface: Provided a public web tool (https://codon.oneshot.ac.cn) supporting 15 major plant species.

4. Key Results

• Model Performance:
  • CodonNAT: Achieved an average prediction accuracy of 66.5% across 15 species (vs. 56.6% for traditional frequency-based methods). It successfully captured phylogenetic relationships (e.g., clustering monocots vs. dicots).
  • CodonEXP: Achieved an average accuracy of 79.3% and AUC of 86.1% across species, outperforming standalone RNA language models.
• Experimental Validation (Tobacco Transient Expression):
  • Tested on 5 benchmark proteins (DsRed2, mCry2Ab, GAT, Infliximab-A/B).
  • DsRed2: Sequences optimized by CodonHa produced fluorescence 1.57x higher than CodonTransformer, 4.32x higher than Genewiz, and 13.58x higher than the traditional BFC method.
  • Large Proteins: Standard optimization failed for large proteins (mCry2Ab, Infliximab-B). The Ha-GC3 variant (incorporating GC3 bias) successfully enabled expression of these proteins.
• Efficiency: CodonHa completed optimization in ~138 seconds (RTX 3090 GPU), compared to ~6,465 seconds for the Genetic Algorithm approach.
• Sequence Characteristics: HalluCodon-generated sequences maintained natural codon usage patterns (CAI ~0.83) while significantly increasing GC3 content (to ~0.30, close to natural genomic levels of 0.31), balancing stability with synthesis feasibility.

5. Significance

• Advancement in Plant Synthetic Biology: HalluCodon provides a robust, data-efficient solution for maximizing protein yields in plant molecular farming and transgenic crop development.
• Paradigm Shift: Moves the field from static, frequency-based optimization to dynamic, context-aware design using foundation models.
• Scalability: The ability to fine-tune pre-trained models on small, species-specific datasets makes high-performance codon optimization accessible for non-model plant species.
• Balanced Optimization: By integrating GC3 enrichment without indiscriminately maximizing GC content, the tool avoids potential issues like gene synthesis difficulty and transcriptional repression due to DNA methylation, offering a practical solution for real-world applications.

In summary, HalluCodon represents a significant leap forward in computational biology for agriculture, combining state-of-the-art language models with efficient generative algorithms to solve the complex problem of species-specific gene expression.