QSAR

QSAR predicts biological activities from chemical structures using ensemble-based machine learning to support drug discovery.

Key Features:

• Ensemble-Based Machine Learning: Builds diversified sets of models across multiple subjects and integrates them via second-level meta-learning to improve predictive performance compared to single-subject ensemble methods such as random forests.
• End-to-End Neural Network Model: Uses an individual classifier that combines 1D Convolutional Neural Networks (1D-CNN) and Recurrent Neural Networks (RNN) to automatically extract sequential features from SMILES representations.
• Meta-Learning Integration: Employs second-level meta-learning to interpret and combine outputs from individual models, enhancing overall prediction accuracy even when individual models have modest standalone performance.

Scientific Applications:

• Drug Discovery: Predicts biological activities from chemical structures to support identification and optimization of potential therapeutic compounds.
• Bioassay Prediction: Evaluated across 19 bioassay datasets, consistently outperforming thirteen individual models and other traditional ensemble approaches.

Methodology:

Construct diverse models that capture different aspects of the data, combine them using second-level meta-learning, and integrate neural network classifiers (1D-CNN + RNN) to extract features from SMILES representations.