Abstract
Molecular representations largely determine the learnability of quantum-chemical properties with machine learning. In order to find the most appropriate way to represent molecules in chemoinformatic studies, a comparative study of nine two-dimensional molecular fingerprints and three RDKit descriptor sets (PHYS, CONF, and PHCO) was conducted in terms of the prediction of five molecular properties by trained predictive machine learning models. The use of RDKit descriptor sets consistently yields more accurate results than hashed fingerprints across properties. Among fingerprints, Layered Fingerprint outperforms for global energy targets (Etot, Eee, Exc), whereas ECFP6 demonstrates better performance for atom-localized (Eatom) and thermodynamic targets (Cp). We further evaluate how the choice of hash function used during fingerprint construction affects representation quality and identify that noncryptographic hashing preserves locality and leads to better and more consistent outcomes than cryptographic hashing (SHA-256). This work provides mechanistic insights into how different molecular representations encode structural and physicochemical information, highlighting the merits and limits of descriptors for learning quantum-chemical properties. This offers practical guidance for selecting molecular representations and hashing strategies in designing and establishing pipelines for the artificial intelligence study of chemistry.