Use NetSolP Online

Commercially Available NetSolP No-Code Web Server

NetSolP: Predicting Protein Solubility and Usability in E. coli with Protein Language Models

NetSolP is a deep learning tool that predicts protein solubility and usability for expression in Escherichia coli directly from amino acid sequence. It was developed by Vineet Thumuluri, Hannah-Marie Martiny, Jose J Almagro Armenteros, Jesper Salomon, Henrik Nielsen and Alexander Rosenberg Johansen and published in Bioinformatics in 2022. The predictor and data are offered through a DTU Health Tech web service, and the authors also release open-source code.

Getting a recombinant protein to express in a soluble, purifiable form is a routine bottleneck in protein production. Experimental screening is slow and costly, so computational filters that rank candidate sequences before they reach the bench are valuable.

The authors argue that earlier solubility predictors were limited by biased training data and weak generalization. They report labeling and His-tag biases in a widely used dataset, and note about 18.6% label disagreement on identical sequences between datasets. NetSolP addresses this by pairing transformer protein language models with new datasets curated using strict sequence-identity partitioning.

Attribute

Details

Task

Predict solubility and usability of proteins expressed in E. coli

Input

Protein amino acid sequence

Model

Ensemble of fine-tuned, dynamically quantized ESM1b protein language models

Faster variant

NetSolP-D, a distilled version (about 3 s per sequence)

Interpretability

Integrated Gradients and conservation-score analysis

Code

Open-source on GitHub; web service at DTU Health Tech

How NetSolP Predicts Protein Solubility from Sequence

Instead of relying on multiple sequence alignments or hand-built features, NetSolP uses the embeddings of pretrained transformer language models.

  1. Language-model embeddings: The authors compared ESM12, ESM-MSA, ESM1b and ProtT5, using a linear classification layer on averaged residue embeddings.

  2. Ensemble and quantization: NetSolP is an ensemble of fine-tuned ESM1b models. Dynamic quantization reduces storage by about a factor of four.

  3. Distillation: NetSolP-D is a distilled model that runs at about 3 s per sequence, compared with about 5 min per sequence for profile-based methods.

  4. Strict data partitioning: Training and test sets were split with a four-phase procedure using a 25% global sequence-identity cutoff, to limit leakage between folds.

  5. Decision threshold: The solubility threshold of 0.69 is the average Youden-index optimum across folds.

Curated Datasets Behind NetSolP

The study uses TargetTrack-derived data (69,420 training and 2,001 test proteins), PSI: Biology (12,216 proteins, reduced to 11,226 for cross-validation), the Price (NESG) set (9,272 usable sequences, with 1,323 as an independent solubility test set and 7,259 for usability) and a Camsol mutation set of 19 proteins and 56 variants.

NetSolP Benchmark Results for Solubility and Usability

On the independent Price solubility test set, NetSolP outperformed the retrained SWI baseline:

Model

Accuracy

Precision

MCC

AUC

NetSolP

0.728

0.773

0.402

0.760

NetSolP-D

0.723

0.769

0.391

0.756

ESM1b-F

0.723

0.764

0.386

0.761

SWI

0.680

0.712

0.269

0.690

In 5-fold cross-validation on PSI: Biology, NetSolP reached 0.70 accuracy and 0.73 AUC, and on the Price usability task it reached 0.65 accuracy and 0.70 AUC. The authors also observed a Spearman correlation of 0.66 between NetSolP and SWI per-amino-acid scores in their interpretability analysis.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

  • A user-friendly graphical interface for setting up and launching experiments

  • A robust API for integration into existing research pipelines

  • An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with NetSolP on Tamarind Bio

  • Construct triage: Rank candidate protein sequences by predicted E. coli solubility before ordering genes or starting expression.

  • Expression planning: Use the usability prediction to flag sequences that may be hard to purify.

  • Large-scale screening: Use the fast distilled model for libraries, since it runs in seconds per sequence.

  • Residue-level insight: Use the authors' interpretability analysis to look at which residues drive a prediction.

How to Use NetSolP on Tamarind Bio

  1. Log in: Sign in at tamarind.bio and open the NetSolP tool.

  2. Provide sequences: Enter or upload the amino acid sequence(s) you want to evaluate.

  3. Choose the model: If exposed, choose between the full ensemble and the faster distilled model.

  4. Run the job: Launch the prediction and wait for it to complete.

  5. Download results: Download the predicted solubility and usability scores.

  6. Interpret scores: The authors use a 0.69 solubility threshold; treat scores near it as uncertain.

  7. Use downstream: Prioritize sequences for cloning and expression, and confirm experimentally.

Things to Keep in Mind

  • Point mutations: NetSolP ranked 37 of 56 Camsol variants correctly (66.1%), so it is weak at detecting small solubility changes from single mutations.

  • Sequence length: Accuracy drops for longer sequences, and residues beyond position 1022 are truncated at prediction time.

  • Scope: Predictions apply to E. coli expression only.

  • Data bias: The authors acknowledge labeling and tag biases in TargetTrack-derived data.

  • Not uniformly best: Some transformer baselines score higher on specific metrics, and experimental validation of importance predictions is left to future work.

Source: Thumuluri V, Martiny H-M, Almagro Armenteros JJ, Salomon J, Nielsen H, Johansen AR. NetSolP: predicting protein solubility in Escherichia coli using language models. Bioinformatics 38(4):941-946, 2022. DOI: 10.1093/bioinformatics/btab801. Code: https://github.com/tvinet/NetSolP-1.0

Supporting 10,000+ scientists around the world,

from leading biotechs, and global biopharma