Research map / Probabilistic and scientific modeling / Molecules and proteins

$\nabla^2$DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials

NeurIPS 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Machine-learned interatomic potentials · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ∇²DFT, a ~16M-conformation DFT (ωB97X-D/def2-SVP) quantum chemistry dataset derived from MOSES drug-like molecules, with splits and a benchmark for neural network potentials.

Problem

Neural network potentials and Hamiltonian-predicting models need large, diverse, accurate quantum chemistry datasets for drug-like molecules, and existing ones lack scale, diversity, or wavefunction/Hamiltonian data.

Approach

Builds a large DFT dataset from MOSES drug-like molecules with multiple clustered conformations per molecule, providing energies, forces, Hamiltonian and overlap matrices, Psi4 wavefunctions, and generalization-tiered test splits (conformation, structure, scaffold) to benchmark NNPs.

Key results

• Dataset contains ~16M conformations for ~2M molecules, 1 to 63 conformations per molecule.
• Benchmarked 8 models predicting energy from 3D structure and 2 models predicting energy and Hamiltonians.
• Total computation required ~60 CPU years; dataset is 15 TB preprocessed and 220 TB full including Psi4 wavefunctions.

Surprising findings

• Several optimization trajectories contain unrealistic energy values where final energy exceeds initial energy.

Limitations

• Limited to small neutral drug-like molecules; unsuitable for charged, open-shell, materials, solvated or protein-ligand systems.
• Forces computed for only about 40% of the dataset.
• DFT inherits functional, basis set and LCAO approximation errors.

Experimental setup

Datasets: MOSES ZINC Clean Leads

Benchmarks: ∇²DFT test splits (conformation, structure, scaffold)

Metrics: energy error force error

Builds on: nablaDFT MOSES

Closest papers by meaning

• Enhancing the Scalability and Applicability of Kohn-Sham Hamiltonians for Molecular Systems (ICLR 2025 · similarity 0.88)
• The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · similarity 0.87)
• EDBench: Large-Scale Electron Density Data for Molecular Modeling (NeurIPS 2025 · similarity 0.87)
• Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.86)
• Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark (NeurIPS 2026 · similarity 0.86)
• Self-Consistency Training for Density-Functional-Theory Hamiltonian Prediction (ICML 2024 · similarity 0.86)