DeepSol FAQ
DeepSol scores 77% accuracy and MCC 0.55, predicts E. coli expression only, works best from 50 to 2,000 residues, and is unreliable on membrane proteins.
These answers cover how accurate DeepSol is, the E. coli expression system and sequence lengths it is trained for, how to handle tags and membrane proteins, and how to read borderline solubility scores.
Frequently asked questions
How accurate is DeepSol?
DeepSol achieves 77% accuracy with a Matthews Correlation Coefficient (MCC) of 0.55 on benchmark datasets for E. coli expression solubility prediction. This represents over 3.5% higher accuracy than PaRSnIP and over 15% higher accuracy than PROSO II. The model was published in Bioinformatics 2018.
What is the underlying technology?
DeepSol uses convolutional neural networks (CNNs) that exploit k-mer structure in protein sequences. The model combines sequence-level features with structural properties derived from amino acid composition to predict solubility.
What organism is this trained for?
DeepSol specifically predicts solubility for proteins expressed in E. coli. Predictions for other expression systems (yeast, insect cells, mammalian) may not be directly applicable.
Can I submit protein sequences with tags?
Yes. Submit the complete sequence as it will be expressed, including any fusion tags (His-tag, GST, MBP, etc.). The model will predict solubility for the full construct.
What about membrane proteins?
Membrane proteins and highly hydrophobic sequences may fall outside the training domain. Predictions for these proteins should be treated with caution; check the confidence score and warnings.
Are results deterministic?
Yes—identical sequences and parameters produce identical outputs.
How big can batches be?
Keep under ~1,000 sequences for optimal turnaround. Larger sets should be chunked; respect rate limits and retry with backoff on 429/503.
Why is confidence low?
Likely out-of-domain sequence characteristics, unusual amino acid composition, or extreme sequence length. Check warnings for specific issues.
What sequence length works best?
The model performs best on sequences between 50-2000 residues. Very short peptides or extremely long proteins may have reduced prediction accuracy.
How should I interpret borderline scores?
Scores near 0.5 indicate uncertain predictions. For these proteins, consider:
- Optimizing expression conditions (temperature, induction level)
- Using solubility-enhancing fusion tags
- Testing multiple constructs with different truncations
- Small-scale expression trials before scale-up
Read next
Was this page helpful?