AI-Driven Molecular Design: Synergizing Deep Generative Models with Evolutionary Optimization

Abstract

Artificial intelligence (AI) is reshaping drug discovery by enabling efficient and precise identification of novel therapeutics. This review examines the synergistic use of deep generative models, such as Long Short-Term Memory (LSTM) networks, Variational Autoencoders (VAEs), and Graph Attention Networks (GATs), together with evolutionary optimization techniques, including genetic algorithms and multi-objective evolutionary strategies. While deep learning architectures excel at capturing complex molecular representations and generating chemically valid compounds, evolutionary algorithms provide complementary strengths in global exploration and multi-objective trade-off optimization. The combination of these two paradigms offers a powerful and complementary toolkit: deep learning provides the capacity to learn rich chemical features and propose innovative scaffolds, whereas evolutionary methods ensure efficient navigation of chemical space and balanced optimization across multiple drug-like criteria. Through comparative analyses, quantitative benchmarks, and illustrative figures, we highlight how integrating generative and evolutionary paradigms can accelerate de novo molecular design, reduce development timelines, and lower costs. We also address technical and ethical challenges. In particular, our ongoing research explores hybrid frameworks that combine variational autoencoders, graph neural predictors, Colibri algorithm and Genetic algorithms with fragment-based crossover, and dynamic multi-objective penalties to further enhance chemical validity, pharmacological relevance, and synthetic accessibility. Future efforts aim to demonstrate that such hybrid frameworks can bridge the gap between theoretical innovation and practical drug development, bringing AI-driven discovery closer to real-world therapeutic breakthroughs