Interpretable Chirality-Aware Graph Neural Networks for QSAR Modeling in Drug Discovery

Explore a conference talk on developing interpretable chirality-aware graph neural networks for quantitative structure-activity relationship modeling in drug discovery. Dive into the limitations of current graph neural networks in capturing molecular chirality and learn about the proposed Molecular-Kernel Graph Neural Network (MolKGNN) approach. Discover how MolKGNN achieves SE(3)-/conformation invariance and interpretability through molecular graph convolution and similarity score propagation. Examine the comprehensive evaluation of MolKGNN across nine datasets featuring high class imbalance, and understand its superiority over other GNNs in computer-aided drug discovery. Gain insights into the interpretability of learned kernels and their alignment with domain knowledge. The talk covers background on message passing schemes, impacts of chirality, intuition from image convolution, similarity score calculation, MolKGNN overview, screening datasets, evaluation metrics, comparison with 3D GNNs, and interpretability results.

- Intro
- Background: Message Passing Scheme
- Graph Neural Network Limitations
- The Impacts of Chirality
- Intuition from Image Convolution
- Similarity Score Calculation
- MolKGNN Overview
- Screening Datasets
- Metrics for Evaluation & Results
- Can MolKGNN Outperform 3D GNNs?
- Intepretability Result
- Conclusion
- Q&A