Computational drug design in the artificial intelligence era: A systematic review of molecular representations, generative architectures, and performance assessment.

Abstract

Generative drug design has emerged as a transformative approach in pharmaceutical research, leveraging deep learning models to create novel molecules with targeted properties. This systematic review analyzes the current landscape of computational approaches across 3 critical dimensions: molecular representation strategies (1-dimensional, 2-dimensional, and 3-dimensional), generative architectural frameworks (including variational autoencoders, generative adversarial networks, reinforcement learning systems, and diffusion models), and evaluation methodologies. We provide a comprehensive categorization of these approaches, critically assessing their relative advantages and limitations. Additionally, we examine the datasets driving development in this field and the metrics employed to evaluate model performance. Through structured analysis of these interconnected components, we identify significant research gaps and propose promising future directions for advancing artificial intelligence-driven drug discovery. This review offers researchers a unified framework for understanding the complex interplay between representation choices, generative mechanisms, and evaluation paradigms in computational drug design. SIGNIFICANCE STATEMENT: This review provides a unique framework for generative drug design by categorizing methods first by drug representation and then by generative model type. This novel taxonomy clarifies which models are best suited for specific molecular data types, offering practical guidance for researchers. We also critically discuss the advantages, disadvantages, key datasets, and evaluation metrics, delivering a comprehensive and actionable resource for the field.