Expression Optimization
Predict protein solubility and guide expression construct design for recombinant protein production.
Expression Optimization
Producing soluble, correctly folded recombinant protein is a prerequisite for structural studies, biochemical assays, and biologic drug development. This workflow uses complementary AI solubility prediction tools to assess expression feasibility, identify aggregation-prone sequences, and guide construct design decisions before committing to time-consuming experimental expression trials.
Overview
Protein insolubility and aggregation are among the most common reasons for failed recombinant expression campaigns. Experimental screening of expression conditions (host organism, temperature, media, fusion tags) is slow and expensive. DeepSol and GraphSol provide complementary predictions: DeepSol uses deep learning on sequence features including amino acid composition, hydrophobicity patterns, and predicted secondary structure, while GraphSol uses graph neural networks that model structure-aware residue interactions. Running both tools provides a consensus solubility assessment with higher confidence than either tool alone.
Tool Chain
- DeepSol -- Sequence-based protein solubility prediction
- GraphSol -- Structure-aware graph neural network solubility prediction
Prerequisites
- Protein sequences in FASTA format or UniProt accessions
- Knowledge of the intended expression host (E. coli, CHO, insect cells)
- Understanding of available fusion tag options (His-tag, MBP, GST, SUMO)
Step-by-Step Guide
Predict solubility with DeepSol
Submit your protein sequences to DeepSol. Review the predicted solubility probability for each construct. DeepSol classifies sequences as soluble (probability above 0.5) or insoluble (below 0.5) and provides a confidence score. Examine the sequence feature analysis to identify regions contributing to poor solubility: hydrophobic patches, charge clusters, and low-complexity regions. Flag any sequence with predicted solubility below 0.4 for construct redesign.
Confirm with GraphSol
Submit the same sequences to GraphSol for an independent assessment. GraphSol incorporates predicted structural features into its analysis, capturing long-range residue interactions that pure sequence models may miss. Compare the GraphSol and DeepSol predictions for each sequence. Sequences where both tools predict high solubility (above 0.6) are strong candidates for direct expression. Sequences where both tools predict low solubility (below 0.4) will likely require significant construct engineering.
Identify problematic regions
For sequences with low predicted solubility, analyze the per-residue contributions from both tools. Look for hydrophobic surface patches (clusters of Leu, Ile, Val, Phe, Trp on the predicted surface), unstructured N-terminal or C-terminal regions (low pLDDT in ESMFold predictions), polyhistidine or polyglutamine stretches, and cysteine-rich regions prone to misfolding in the reducing cytoplasm of E. coli.
Design improved constructs
Based on the identified problematic regions, design alternative expression constructs. Common strategies include truncating disordered termini that do not contribute to function, adding solubility-enhancing fusion tags (MBP or SUMO for E. coli; Fc for mammalian cells), replacing surface-exposed hydrophobic residues with polar alternatives (consensus design), codon-optimizing for the target expression host, and splitting multi-domain proteins into individually expressible domains.
Predict solubility of redesigned constructs
Run the modified constructs back through DeepSol and GraphSol to verify improvement. A successful redesign should show a solubility probability increase of at least 0.15 from the original sequence. If the improvement is insufficient, consider further modifications or alternative expression strategies such as cell-free expression systems, refolding from inclusion bodies, or switching to a eukaryotic host.
Prioritize constructs for experimental testing
Rank all construct variants by consensus solubility prediction, predicted structural integrity (ESMFold pLDDT for key functional regions), and practical considerations (expression host availability, purification tag compatibility, downstream assay requirements). Select 3-5 top-ranked constructs for parallel experimental expression trials to hedge against prediction uncertainty.
Expected Outputs
- DeepSol solubility probability scores with sequence feature analysis
- GraphSol solubility predictions with structure-aware residue contributions
- Consensus solubility assessment highlighting agreement and disagreement
- Problematic region identification with specific residue annotations
- Redesigned construct sequences with predicted solubility improvements
- Prioritized construct list for experimental expression testing
Tips and Best Practices
- Always run both DeepSol and GraphSol. Agreement between independent models is the strongest predictor of experimental outcome. When the tools disagree by more than 0.3 probability, treat the prediction as uncertain and plan experimental validation.
- For E. coli expression, treat agreement above 0.7 from both tools as a strong go signal and agreement below 0.3 as a signal to add a fusion tag before attempting direct expression.
- MBP (maltose-binding protein) fusion is the most effective solubility tag for E. coli expression. Include a TEV protease site for tag removal if the native protein is needed.
- Consider expressing protein domains separately when the full-length prediction is poor but individual domain predictions are acceptable. Domain boundaries can be identified from ESMFold confidence plots.
- Codon optimization alone rarely rescues an insoluble protein. Focus on sequence-level and construct-level changes first, then optimize codons for expression level.
- Track the correlation between predicted and observed solubility for your protein targets. Over time, this builds a calibration dataset that improves decision confidence for future targets.
Was this page helpful?