Enhancing AI-aided drug design with Uni-QSAR (on behalf of Dr. Cui)
Date:
The application of AI/ML in drug discovery faces critical challenges in data scarcity scenarios, particularly in deriving robust quantitative structure-activity relationships (QSAR) with limited molecular annotations. We present Uni-QSAR, an automated framework that overcomes these limitations through multidimensional molecular representation learning. Our approach innovatively integrates 1D sequence patterns, 2D topological graphs, and 3D conformer geometry into a unified embedding space, enhanced by pretraining on large-scale unlabeled molecular data. Without manual parameter tuning, Uni-QSAR achieves state-of-the-art performance on major benchmarks: outperforming existing methods in 14/22 tasks (Therapeutic Data Commons) and 26/30 tasks (MoleculeACE). Notably, we demonstrate through Bindingnet analysis that 3D conformer-aware modeling significantly improves binding activity prediction accuracy. This fully automated system provides researchers with an efficient tool for predictive molecular profiling, substantially advancing AI-driven drug discovery pipelines by reducing dependency on labeled data while maintaining prediction robustness.