De novo drug design is a cutting-edge approach that aims to generate novel chemical compounds from scratch, offering great potential for accelerating therapeutic development in various medical domains. However, designing effective and safe drugs de novo remains a complex challenge due to the need to satisfy multiple constraints, including drug-likeness, pharmacokinetics, and target specificity. Traditional computational methods, while useful, often struggle to fully explore the vast chemical space and optimize molecules for multiple objectives. In recent years, the integration of artificial intelligence (AI), particularly deep learning (DL), has transformed the landscape of de novo drug design. DL models have demonstrated the ability to learn complex molecular representations, predict bioactivity, and generate novel drug-like compounds with desirable properties. This review provides a concise overview of recent advances in deep learning-driven de novo drug design, highlighting the key architectures and methodologies employed in this domain. Special attention is given to the chemical property optimization process, a critical step for ensuring the therapeutic relevance of generated molecules. Furthermore, we explore case studies across specific disease areas to illustrate how deep learning frameworks have been applied to tailor molecular generation toward medical needs. Finally, current challenges and future research directions are discussed to inform the development of more efficient and clinically relevant AI-based drug discovery pipelines.