SolvNet 1.0.20

Protein solubility prediction toolkit

SolvNet

Protein Solubility Prediction Toolkit

SolvNet is a Python toolkit for predicting protein solubility from amino acid sequences. It provides easy-to-use functions to generate embeddings and predict solubility with confidence estimates.

---

Installation

Create environment:

conda create --name solvnet
conda activate solvnet

Install package

pip install SolvNet

Step 0: Prepare Data

Prepare your dataset or load a test dataset from pbdsolv.

from SolvNet import load_pdb
df = load_pdb()

The dataset should be a pandas DataFrame with the following columns:

Column	Description
name	Protein name
aa_seq	Amino acid sequence
label	Binary solubility label (0: insoluble, 1: soluble)

Step 1: Generate ESMC Embeddings

Generate ESMC embeddings for the protein sequences:

from SolvNet import esmc_embeddings
data = esmc_embeddings(df_test)

Parameters:

• df_test: Dataframe from step 0.

Step 2: Predict Solubility

Use the generated embeddings to predict protein solubility:

from SolvNet import Predict

labels, preds, confs, probs = Predict(
    data=data,
    model = "pdb", # model in ["pdb", "plm"]
    percentage=0.85,
    batch_size=64
)

Parameters:

• data: ESMC embeddings.
• model: Must be pdb or plm, which indicates the dataset the model was trained on Default is pdb
• percentage: The top percentage of most confident predictions to return (default 0.85).
• batch_size: Batch size for prediction (default 64).

Returns:

• labels: True labels for the top percentage% most confident proteins.
• preds: Predicted labels for the top percentage%.
• confs: Confidence scores for predictions.
• probs: Predicted probabilities.

Example Workflow for Inference

from SolvNet import *

# Step 0: Load pdb dataset:
df = load_pdb()

# Step 1: Generate embeddings
data = esmc_embeddings(df)

# Step 2: Make predictions
labels, preds, confs, probs = Predict(
    data=data,
    percentage=0.85,
    batch_size=64
)

Training

Step 0: Prepare df_train, df_val, and df_test

from SolvNet import *

top_percent = 0.95
noise_rate = 1-top_percent

TRAIN, VAL, TEST = esmc_embeddings(df_train), esmc_embeddings(df_val), esmc_embeddings(df_test)
TRAIN, VAL, TEST = list(TRAIN.values()), list(VAL.values()), list(TEST.values())

best_model, last_model, metrics = train_dividemix(TRAIN, VAL, TEST, batch_size=64, lr=1e-4, epochs=12, warmup_epochs=5, noise_rate=noise_rate, alpha=0.85) # alpha: hyperparameter for mixup augmentation

---

License

MIT License © 2025 Thao Nguyen et al.