Protein Expression Rescue

Coupling inverse folding sequence generators with an in-house LLM-based expression prediction model, we rescue non- and low-expressing proteins of interest by proposing variant sequences that reliably express at 10-100x the level of the parent sequence.

Problem

A major bottleneck in protein engineering and therapeutic development is poor expression in production systems. Many proteins of interest:

• Fail to express in bacterial or mammalian hosts
• Express at levels too low for purification
• Form insoluble aggregates
• Undergo degradation

Traditional approaches require extensive trial-and-error mutagenesis, which is time-consuming and expensive.

Approach

We combine two powerful AI tools:

1. Inverse Folding Models: Generate sequence variants that maintain the desired 3D structure
2. Expression Prediction LLM: Predict expression levels from sequence alone

Our pipeline:

1. Input: Low/non-expressing protein structure or sequence
2. Generate diverse sequence variants using inverse folding (e.g., ProteinMPNN, ESM-IF)
3. Score variants with our expression prediction model
4. Select top candidates balancing structure preservation and high expression
5. Validate experimentally

Expression Prediction Model

Our in-house LLM for expression prediction:

• Training Data: Millions of expression measurements across diverse proteins
• Architecture: Transformer-based model leveraging protein language model embeddings
• Features: Considers sequence properties, codon usage, structural features
• Output: Quantitative expression level predictions with uncertainty estimates

Results

Across multiple challenging cases:

• 10-100x expression improvement over parent sequences
• High success rate: >80% of top predictions express well
• Structure preservation: Variants maintain target fold
• Fast turnaround: Days instead of months for optimization

Applications

• Therapeutic Development: Rescue antibodies, enzymes, and novel scaffolds
• Structural Biology: Enable crystallography/cryo-EM studies requiring high protein yields
• Industrial Enzymes: Improve production economics
• Synthetic Biology: Design highly expressible parts for genetic circuits

Technical Advantages

• AI-Driven: Leverage recent advances in protein language models
• Structure-Aware: Maintain functional architecture while optimizing expression
• Uncertainty Quantification: Know confidence in predictions
• Iterative Refinement: Can be applied multiple times for further optimization

Impact

This approach dramatically reduces the time and cost of protein engineering projects by:

• Eliminating months of screening poorly expressing variants
• Reducing need for expensive mammalian expression systems
• Enabling previously intractable projects
• Accelerating therapeutic development timelines