GEx Predictor

GEx Predictor is a tool for predicting gene expression based on chemical molecules represented as SMILES. It uses a trained neural network to generate predictions from global molecular signatures produced with Signaturizer.

Features

• Standardization of SMILES using RDKit.
• Generation of global molecular signatures (GLOBAL Signature) with Signaturizer.
• Gene expression prediction using a deep learning model (GenomicExpressionNet2).
• Compatible with GPU and CPU via PyTorch.

Data and Model Training

The predictive model in GEx Predictor has been trained using data from the LINCS (Library of Integrated Network-based Cellular Signatures) project, which provides large-scale gene expression profiles in response to chemical perturbations.

To enrich the dataset and cover molecules not directly available in LINCS, we leveraged the Chemical Checker (CC) resource developed by the Structural Bioinformatics and Network Biology (SBNB) group at IRB Barcelona. The workflow was as follows:

1. LINCS data: Experimental gene expression signatures were used as the foundation (~20,000 molecules).
2. Chemical Checker molecules: Additional molecules from the CC were included to expand the chemical space.
3. KNN imputation: For molecules in the Chemical Checker that are close to LINCS compounds in signature space, we applied k-nearest neighbors (KNN) imputation to estimate their gene expression profiles.

As a result, the final training dataset combines:

• ~20,000 experimental LINCS profiles, and
• ~80,000 imputed gene expression profiles for chemically similar molecules (from the CC).

In total, the dataset contains approximately 100,000 molecules.

Model Training Strategy

• To ensure robust evaluation and avoid overfitting to specific chemotypes, we used a scaffold split: molecules were partitioned based on their chemical scaffolds rather than random sampling.
• The model was trained and validated using 5-fold cross-validation, where each fold uses a different subset of scaffolds for testing and the remaining scaffolds for training/validation.
• This strategy guarantees that the model is tested on novel scaffolds not seen during training, providing a more realistic estimate of generalization performance.

GLOBAL Molecular Signature

This project uses the GLOBAL molecular signature generated by Signaturizer. This signature represents the chemical and bioactive properties of a molecule in a comprehensive way, as a 3,200-dimensional vector.

Key details

• GLOBAL signatures combine multiple bioactivity spaces, capturing a complete representation of the molecule.
• Each signature is a high-dimensional vector summarizing chemical and biological information.
• Signatures are generated via Siamese neural networks trained with a triplet approach (anchor, positive, negative), allowing bioactively similar molecules to have close representations in vector space.
• These representations are useful for predicting biological properties, comparing compounds, and analyzing chemical similarity.

Workflow in the predictor

1. The molecule is standardized (clean-up, largest fragment, uncharging, tautomerization).
2. The GLOBAL signature of the molecule is generated.
3. The signature is converted to a tensor and fed into the trained neural network to obtain gene expression predictions.

Installation

1. Clone the repository:

git clone https://github.com/AdriaVico/GEx_Predictor.git
cd GEx_Predictor

Usage

from gex_predictor import GEx_Predictor

# Initialize the predictor
predictor = GEx_Predictor()

# Prediction from SMILES — input must be a list of SMILES strings
smiles = ["CCO", "c1ccccc1"]
prediction = predictor.predict(smiles, input_type="SMILES")
print(prediction)  # numpy array of shape (n_molecules, 1035)

# Return a DataFrame with LINCS landmark gene names as columns
df = predictor.predict(smiles, input_type="SMILES", as_dataframe=True)
print(df)

# Get the ordered list of predicted gene names
genes = predictor.get_genes()

Project Structure

GEx_Predictor/ ├─ model/ │ └─ fold_4.pt # Trained model ├─ predictor.py # GEx_Predictor class ├─ requirements.txt # Dependencies └─ README.md

Requirements

• Python >= 3.8
• PyTorch
• RDKit
• Signaturizer

Contributing

Contributions are welcome for improving molecule standardization, model optimization, and adding new prediction features.