tCNNS

tCNNS predicts phenotypic drug responses in cancer cell lines by using twin convolutional neural networks to model interactions between drugs (SMILES) and genetic feature vectors.

Key Features:

• Dual Convolutional Networks: Two separate CNNs extract features from drug representations in SMILES format and from genetic feature vectors of cancer cell lines.
• Integration and Prediction: A fully connected network integrates the extracted features to predict drug–cell line interactions (phenotypic response).
• Performance Metrics: Reported results include mean R² of 0.826, top quartile R² of 0.831, mean Pearson correlation (R_p) of 0.909, and top quartile R_p of 0.912.
• Data Efficiency: Maintains high accuracy with limited but high-quality data and requires fewer features for cancer cell lines.
• Outlier Management: Effectively addresses outliers in other feature spaces to enhance robustness.

Scientific Applications:

• Phenotypic Screening: Predicts drug responses to support identification of therapeutic candidates and drug repurposing.
• Genomic Data Analysis: Uses genetic data from cancer cell lines and aligns predictions with the Genomics of Drug Sensitivity in Cancer (GDSC) database.

Methodology:

Two CNNs process SMILES and genetic feature vectors, a fully connected network integrates extracted features to predict interactions, datasets are divided by drug–cell line interaction pairs for training and testing, and performance is evaluated using R² and Pearson correlation (R_p); performance decreases when training/testing splits are made exclusively by drugs or exclusively by cell lines.