Molecular Representations for Machine Learning

Machine-learning systems need a numerical representation of molecules. SMILES serializes a molecular graph as a character string. It is compact and convenient for sequence models, but one molecular graph can have multiple valid SMILES traversals. Canonicalization can provide a consistent form, yet the string order is not itself a physical property. Tokenization and chemical parsing matter. Fingerprints convert molecular substructures or features into fixed-length vectors. Circular fingerprints encode atom neighborhoods; path-based fingerprints encode graph fragments. They are efficient for similarity search and classical QSAR, but folding features into a fixed bit vector can create collisions, and fingerprints may discard some stereochemical or spatial detail depending on settings. A molecular graph represents atoms as nodes and bonds as edges, often with attributes such as element, charge, aromaticity, and bond type. Graph neural networks learn representations by passing information across bonds. This preserves connectivity directly but requires choices about atom and bond features, message-passing depth, and pooling. A graph model does not automatically know three-dimensional conformations unless geometry is included. Three-dimensional representations add atom coordinates, distances, angles, or conformer ensembles. They can represent shape and spatial interactions but depend on conformer generation, protonation, stereochemistry, alignment, and coordinate quality. A single conformer may miss flexibility. Structural models need physically and chemically reasonable input preparation. Representation choice should follow the endpoint and data scale. A fingerprint baseline can be strong and easy to validate, while graph or 3D models may capture richer structure at greater cost. Standardize structures consistently, define how salts and tautomers are handled, and split by scaffold when testing chemical generalization. Avoid assuming that one representation captures every property or that representation complexity guarantees better prediction.