DeepSol on the platform
Predict protein solubility for E. coli expression from an amino acid sequence.
DeepSol predicts protein solubility from sequence using a convolutional neural network, flagging whether a protein is likely to express solubly. Reach for it when a solubility call gates the next step — screening expression constructs or triaging candidates before wet-lab work. It predicts solubility by a different method than GraphSol, so run both and treat agreement as a consensus. See what you can achieve for the full set.
What DeepSol does
DeepSol is a deep learning-based protein solubility predictor using convolutional neural networks (CNNs). Published in Bioinformatics, it exploits k-mer structure and additional sequence/structural features to predict whether proteins will be soluble when expressed in E. coli.
Citations
Khurana S, Rawi R, Kunji K, Chuang GY, Bensmail H, Mall R. DeepSol: a deep learning framework for sequence-based protein solubility prediction. Bioinformatics. 2018 Aug 1;34(15):2605-2613. doi: 10.1093/bioinformatics/bty166
Performance benchmarks
DeepSol outperformed all known sequence-based solubility prediction methods:
- Accuracy: 0.77
- Matthew's correlation coefficient (MCC): 0.55
- At least 3.5% more accurate than PaRSnIP
- At least 15% more accurate than PROSO II
DeepSol is freely available at https://zenodo.org/record/1162886.
Webapp integration and capabilities
Input in .bloom*
- Amino acid sequence: Paste standard 1-letter-code sequence directly, or fetch by UniProt accession ID (e.g.
P00533). - Model: Choose
deepsol1(default),deepsol2, ordeepsol3. - SCRATCH-1D features: Enabled by default; adds structural features to the CNN input alongside the k-mer sequence features.
Prediction outputs
- Solubility score: 0–1, where a score above 0.5 indicates likely soluble expression.
- Solubility class:
solubleorinsoluble, from the 0.5 threshold. - Confidence score: model certainty for the call.
Model characteristics and limitations
- Sequence length: 20–2000 residues (see inputs & parameters for the recommended 50–2000 range).
- Standard residues only: non-standard amino acids (X, B, Z, J) are rejected.
- Applicability: trained on E. coli expression data — accuracy for other expression hosts hasn't been validated by the platform.
- DeepSol predicts by CNN sequence features; GraphSol predicts from a structure-aware graph built on predicted contact maps. The two use different signals, so treat agreement between them as a consensus rather than relying on either alone.
Chat Usage
Example Prompts
- "Predict solubility for this protein sequence using DeepSol"
- "Is this protein likely to be soluble? Run DeepSol"
Workflow Chaining
- Feeds into: expression-construct triage, GraphSol consensus check
- Accepts from: UniProt (sequences), ESMFold (predicted structures)
Read next
GraphSol
Predict protein solubility from sequence with a graph convolutional network.
DeepSol inputs and parameters
Protein sequence requirements and parameters for DeepSol solubility prediction.
Which tool?
Match a scientific goal to the tool that answers it.
Workflows
Guided pipelines and copy-and-adapt prompt recipes for common goals.
Was this page helpful?
Troubleshooting DeepPurpose
Fix DeepPurpose installs, DGL and RDKit imports, CUDA device errors, out-of-memory training runs, SMILES parsing failures and diverging loss or NaNs.
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.