---
title: DeepSol results and interpretation
description: >-
  Understand protein solubility scores and prediction confidence from DeepSol
  CNN.
---

You get back a solubility probability (0–1) and a soluble/insoluble call for each sequence. That score sets your next move: a high score prioritizes a protein for direct expression, while a low one flags likely inclusion-body formation, so you can plan solubility tags, refolding, or an alternative host before running the trial.

## Output fields

- **`id`**: your submitted identifier from FASTA header or CSV (if provided).
- **`sequence`**: the input amino acid sequence processed by the model.
- **`solubility_score`**: predicted probability of soluble E. coli expression (0-1 scale).
- **`prediction`**: binary classification (`soluble` or `insoluble`) based on 0.5 threshold.
- **`confidence`**: model certainty derived from CNN output layer.
- **`warnings`**: array of strings with validation or domain-fit notes.

## Interpreting solubility scores

The solubility score represents the probability that a protein will be expressed in soluble form in E. coli:

- **0.7-1.0 (High):** Strong likelihood of soluble expression; prioritize for direct expression trials.
- **0.5-0.7 (Medium):** Moderate solubility expected; consider optimizing expression conditions or using solubility tags.
- **0.3-0.5 (Low):** Likely to form inclusion bodies; plan for refolding protocols or alternative expression systems.
- **0.0-0.3 (Very Low):** High probability of insoluble expression; consider fusion partners, truncations, or alternative hosts.

## Model performance context

DeepSol achieves 77% accuracy with a Matthews Correlation Coefficient (MCC) of 0.55 on benchmark datasets. This represents:
- Over 3.5% higher accuracy than PaRSnIP
- Over 15% higher accuracy than PROSO II

The CNN architecture exploits k-mer structure in protein sequences and combines sequence-level features with structural properties derived from the amino acid composition.

## Calibration and limits

- Model trained and validated on E. coli expression data; predictions are specific to bacterial expression systems.
- Predictions for proteins far from the training domain (e.g., membrane proteins, highly disordered sequences) may be less reliable.
- The model predicts intrinsic solubility tendency; actual expression outcomes also depend on conditions, strain, and protocol.
- For critical decisions, validate predictions experimentally with small-scale expression trials.

## Quality checks

- Confirm `warnings` is empty before downstream use; otherwise flag for manual review.
- Track distribution of scores across your protein set to identify potential batch effects.
- Compare predictions against any known soluble/insoluble proteins in your dataset as internal controls.
- Consider the biological context: some proteins may require cofactors or partners for proper folding.

## Canvas Visualization

When a DeepSol job completes, a result card appears on the canvas.

- **Card size**: 350x300px
- **Border color**: Orange (#FF9800)
- **Visualization**: Solubility prediction summary

### Tier 2 Metrics

- Solubility score
- Confidence percentage
- LogP
- Molecular weight
- HBD / HBA counts

## Read next