DEEPScreen

DEEPScreen predicts drug-target interactions by applying deep convolutional neural networks to 2-D structural representations of compounds for virtual screening in drug discovery.

Key Features:

• Deep Convolutional Neural Networks: Employs deep CNNs to analyze compound images derived from 2-D structural representations and automatically extract features for DTI prediction.
• 2-D Structural Representations: Uses 2-D compound images instead of conventional molecular descriptors to learn complex structural information relevant to interactions.
• Training Data: Models were trained on curated bioactivity data covering 704 target proteins.
• Hyper-parameter Optimization: Model performance was refined through rigorous hyper-parameter optimization tests.
• Benchmarking: Performance was compared against state-of-the-art methods on multiple benchmark datasets.
• Computational Validation: Selected novel predictions were validated using molecular docking analysis and literature-based verification.
• Experimental Validation: Predicted JAK proteins as new targets for cladribine and confirmed the prediction in vitro by observing effects on STAT3 phosphorylation in cancer cells.

Scientific Applications:

• Virtual Screening: Prioritizes compound-target pairs in large-scale in silico screening campaigns for drug discovery.
• Drug Repurposing: Generates hypotheses for new targets of existing drugs, exemplified by the cladribine–JAK prediction.
• Chemogenomic Screening: Explores chemogenomic space to identify candidate drug-target interactions for experimental follow-up.

Methodology:

Analyzes 2-D compound images with deep convolutional neural networks; trains models on curated bioactivity data across 704 target proteins; applies rigorous hyper-parameter optimization; benchmarks against state-of-the-art methods on multiple datasets; and performs molecular docking analysis and literature-based verification for computational validation.