This chapter delves into the characterization of small-molecule compounds, bridging traditional knowledge-based methods and advanced deep learning approaches. Traditional techniques rely on expert-designed molecular descriptors and fingerprints that capture molecular structures, physicochemical properties, and biological activities. Widely used in drug development, these methods are limited by their dependency on expert intuition. Recent advancements in deep learning enable models to automatically extract features from primitive molecular representations such as strings (e.g., Simplified Molecular Input Line Entry System (SMILES)), graphs, and images. Graph-based representations treat molecules as nodes (atoms) and edges (bonds), providing intuitive depictions and leveraging graph neural networks for property prediction. String and image-based methods offer compact, versatile representations, allowing deep learning models to learn molecular features directly. These approaches reduce reliance on domain-specific knowledge while uncovering new molecular property prediction possibilities. However, challenges such as data requirements and overfitting remain. The chapter underscores the evolution, applications, and future potential of molecular representation techniques, driving innovation in cheminformatics and drug discovery.