PC77M from PubChemQC

This repository provides a curated collection of molecular datasets derived from the original PubChemQC database. The original PubChemQC database contains approximately 86 million molecules and is available at: https://nakatamaho.riken.jp/pubchemqc.riken.jp/ During the curation process, all entries with invalid or non-parsable SMILES representations were removed, while all molecules with valid SMILES strings were preserved. This procedure resulted in a curated dataset containing approximately **77 million molecules**, referred to as **PC77M**. Each molecule in the dataset includes the following information: * SMILES representation; * HOMO (Highest Occupied Molecular Orbital) energy; * LUMO (Lowest Unoccupied Molecular Orbital) energy. In addition to the complete PubChemQC77M dataset, several benchmark subsets of different sizes were generated to facilitate reproducible experiments and enable the evaluation of machine learning models under different data availability scenarios. The available subsets contain: * 50K molecules * 100K molecules * 200K molecules * 500K molecules * 1M molecules * 2M molecules * 5M molecules * 10M molecules * 20M molecules ## Dataset Generation The complete PubChemQC77M dataset was randomly partitioned into **90% training**, **5% validation**, and **5% test** sets. All smaller datasets were generated hierarchically from the training portion of the immediately larger dataset. Specifically, the 20M dataset was sampled from the 77M training set; the 10M dataset was sampled from the 20M training set; the 5M dataset was sampled from the 10M training set, and so on until the 50K dataset was obtained. For every dataset size, a new **90%/5%/5% split** (training/validation/test) was performed independently. This hierarchical construction guarantees that: * training, validation, and test sets are mutually exclusive within each dataset; * no data leakage occurs between the training, validation, and test partitions; * smaller datasets are strict subsets of the larger ones, enabling fair and consistent scalability studies. These datasets are intended for benchmarking machine learning methods for molecular property prediction, representation learning, generative models, and related applications in computational chemistry and materials science.