TAL-GA: A task-adaptive and lightweight generative architecture for multi-property molecular design in drug discovery

Deep learning-based generative models have attracted significant attention for property-driven molecular design and functional molecule discovery, especially in novel drug development. However, most state-of-the-art inverse molecular design models require computationally intensive retraining when optimization objectives change, limiting their practical adaptability. As such, we propose a task-adaptive and lightweight generative architecture (TAL-GA) for multi-property molecular design, which adopts a decoupled framework, integrateing a dynamic molecular generation module with a static property prediction module, leveraging a Variational Auto-Encoder, Deep Pyramid Convolutional Neural Network, and Swarm Intelligence-based optimization strategy. By embedding pretrained property predictors directly into a gradient-free optimization loop, TAL-GA eliminates the need for retraining the generative model when design objectives are modified. Evaluated through two drug discovery case studies, TAL-GA demonstrates superior performance in generating valid, diverse, and novel drug-like molecules surpassing reference models. By combining computational efficiency with high adaptability to evolving design criteria, TAL-GA offers a transformative approach to accelerating drug development, particularly for challenging small-molecule and target-specific drug design tasks. This retraining-free, swarm-based architecture offers a task-adaptive, efficient and robust paradigm for addressing complex molecular inverse design challenges in drug discovery.