[Paper Review] Alchemy: A Quantum Chemistry Dataset for Benchmarking AI Models

Guangyong Chen, Pengfei Chen+10 more|arXiv (Cornell University)|Jun 22, 2019

Machine Learning in Materials ScienceMaterials Science37 references64 citations

TL;DR

Alchemy introduces a large quantum chemistry dataset of 119,487 organic molecules with 12 quantum mechanical properties to benchmark ML models, and benchmarks various GNN architectures on it.

ABSTRACT

We introduce a new molecular dataset, named Alchemy, for developing machine learning models useful in chemistry and material science. As of June 20th 2019, the dataset comprises of 12 quantum mechanical properties of 119,487 organic molecules with up to 14 heavy atoms, sampled from the GDB MedChem database. The Alchemy dataset expands the volume and diversity of existing molecular datasets. Our extensive benchmarks of the state-of-the-art graph neural network models on Alchemy clearly manifest the usefulness of new data in validating and developing machine learning models for chemistry and material science. We further launch a contest to attract attentions from researchers in the related fields. More details can be found on the contest website \footnote{https://alchemy.tencent.com}. At the time of benchamrking experiment, we have generated 119,487 molecules in our Alchemy dataset. More molecular samples are generated since then. Hence, we provide a list of molecules used in the reported benchmarks.

Motivation & Objective

• Motivate the need for a larger, more diverse quantum-chemistry dataset beyond MoleculeNet and QM-series.
• Create a medicinal-chemistry–focused QM dataset with broader atomic diversity and size.
• Provide benchmarks of state-of-the-art graph neural networks on quantum property prediction tasks.

Proposed method

• Compile a new molecular dataset (Alchemy) with 12 quantum mechanical properties for 119,487 organic molecules from GDB MedChem.
• Compute properties using PySCF at the DFT B3LYP/6-31G(2df,p) level.
• Represent molecules as graphs and apply graph neural networks to predict quantum properties.
• Benchmark multiple GNN architectures (e.g., GCN, GAT, RGCN, GGNN, MPNN, LanczosNet, GIN).
• Discuss data generation workflow and runtime characteristics for processing molecules.

Experimental results

Research questions

• RQ1 How does ML-based molecular property prediction perform on a diverse QM dataset beyond QM9/MoleculeNet?
• RQ2 Which graph neural network architectures yield the best accuracy for quantum mechanical property prediction on Alchemy?
• RQ3 Does increased molecular size and atom-type diversity improve generalization and transferability of ML models in quantum chemistry?
• RQ4 What are the practical considerations (computational cost, data generation) for large-scale quantum chemistry datasets?

Key findings

• Alchemy contains 119,487 molecules with 12 quantum mechanical properties.
• Molecules have up to 14 heavy atoms (C, N, O, F, S, Cl) and are sourced from the GDB MedChem subset.
• PySCF with B3LYP/6-31G(2df,p) was used to compute the properties.
• State-of-the-art GNN models were implemented and benchmarked on Alchemy.
• The dataset expansion aims to aid evaluation, benchmarking, and development of ML methods for chemistry and materials science.