Deep Learning Survey Benchmarks Molecular Property Prediction Methods

A thorough investigation explores deep learning techniques aimed at forecasting molecular properties, linking molecular structure to both physicochemical and biological behaviors. The analysis focuses on four unique paradigms: Quantum, Descriptor Machine Learning, Geometric Deep Learning, and Foundation Models. This research proposes a cohesive taxonomy that connects molecular representations, model architectures, and their interdisciplinary applications. Benchmark assessments draw from popular datasets alongside those that reflect industry insights, encompassing quantum, physicochemical, physiological, and biophysical areas. The study critically evaluates current practices in data curation, splitting methods, and evaluation protocols. It highlights several challenges, such as inconsistent stereochemistry and issues with reproducibility, urging the need for advancements in the discipline. Insights from quantum chemistry, cheminformatics, and deep learning are integrated into this work.