QCell: Comprehensive Quantum-Mechanical Dataset Spanning Diverse Biomolecular Fragments

Recent advances in machine learning force fields (MLFFs) are revolutionizing molecular simulations by bridging the gap between quantum-mechanical (QM) accuracy and the computational efficiency of mechanistic potentials. However, the development of reliable MLFFs for biomolecular systems remains constrained by the scarcity of high-quality, chemically diverse QM datasets that span all of the major classes of biomolecules expressed in living cells. Crucially, such a comprehensive dataset must be computed using non-empirical or minimally empirical approximations to solving the Schrödinger equation. To address these limitations, we introduce the QCell dataset – a curated collection of 525k new QM calculations for biomolecular fragments encompassing carbohydrates, nucleic acids, lipids, dimers, and ion clusters. QCell complements existing datasets, bringing the total number of available data points to 41 million molecular systems, all calculated using hybrid density functional theory with nonlocal many-body dispersion interactions, as captured by the PBE0+MBD(-NL) level of quantum mechanics. The QCell dataset therefore provides a valuable resource for training next-generation MLFFs capable of modeling the intricate interactions that govern biomolecular dynamics beyond small molecules and proteins.

💡 Research Summary

The manuscript introduces QCell, a large‑scale quantum‑mechanical dataset specifically designed to fill the long‑standing gap in high‑quality QM reference data for the three major classes of biomolecules—nucleic acids, lipids, and carbohydrates—that together account for roughly 40 % of cellular mass. While existing datasets such as QM9, ANI‑1, MD17, QCML, and GEMS provide extensive coverage of small organic molecules and protein fragments, they lack systematic representation of DNA/RNA fragments, phospholipid assemblies, and saccharide motifs. QCell contributes 525 000 newly computed fragments ranging from 2 to 402 atoms, all evaluated at the hybrid density‑functional level PBE0 combined with many‑body dispersion (MBD) and its non‑local variant (MBD‑NL). This level is non‑empirical, offering a reliable description of non‑covalent interactions (π‑π stacking, sterol‑lipid contacts, glycosidic linkages) that are crucial for realistic biomolecular modeling.

The authors describe a five‑step workflow: (1) construction of a curated library of building blocks (DNA/RNA trimers, phospholipid head‑group and tail fragments, monosaccharides and disaccharides, ion‑water clusters); (2) extensive conformational sampling using classical MD (OL21 for nucleic acids, Lipid21 for membranes, TIP3P water) and the CREST conformer generator for sugars; (3) selection of representative fragments based on geometric criteria (distance thresholds for dimers/trimers) and chemical relevance (e.g., base‑pair stacking, cholesterol‑lipid interactions, glycosylation sites); (4) rapid pre‑optimization with semi‑empirical DFTB+MBD to eliminate high‑energy clashes; and (5) high‑accuracy single‑point calculations with the all‑electron FHI‑aims code using the def2‑TZVPP basis set where appropriate. The resulting HDF5 files store atomic numbers, coordinates, total and formation energies, Kohn‑Sham eigenvalues, exchange‑correlation contributions, Hartree‑Fock, kinetic, electrostatic, van‑der‑Waals, and dispersion‑specific parameters (C₆, a₀).

Dataset composition is detailed in Table I. Nucleic‑acid entries include 5 333 solvated DNA duplexes (2 bp) and 9 534 solvated DNA duplexes (3 bp), plus 19 971 gas‑phase RNA fragments. Lipid entries comprise 12 000 fatty‑acid clusters (1‑3 mers) and 4 000 cholesterol‑containing clusters. Carbohydrate entries consist of 59 156 disaccharides covering all α/β anomers and linkages, and 14 931 glycosidic‑linkage fragments. Ion and water entries add 25 000 solvated ion clusters and 5 000 water clusters of varying sizes. Together with the 370 956 DES370K dimers and the 525 881 new QCell fragments, the total reaches 41 036 729 QM points spanning 82 elements (H, C, N, O, P, S, Na, K, Cl, Mg, Ca, etc.).

A key methodological choice is the consistent use of PBE0+MBD(-NL) across QCell and the previously published datasets (QCML, QM7‑X, AQM, GEMS, SPICE). This uniformity enables straightforward merging into a single training set without the need for energy‑scale corrections or re‑parameterization, facilitating the development of truly transferable machine‑learning force fields. The authors argue that hybrid functionals with many‑body dispersion capture both short‑range covalent chemistry and long‑range van‑der‑Waals forces with an accuracy comparable to coupled‑cluster reference data for systems up to ~150 atoms, while remaining computationally tractable for the ~525 k calculations required.

The paper emphasizes the practical impact of QCell: it provides the missing high‑level QM reference for membrane simulations, nucleic‑acid dynamics, and carbohydrate recognition, thereby allowing MLFFs to be trained on data that reflect the actual physicochemical environments of cellular components. The open‑source release of the dataset and the accompanying workflow scripts on GitHub further encourages community adoption, reproducibility, and future expansion (e.g., inclusion of post‑translational modifications, larger oligomers, or explicit solvent models).

In conclusion, QCell represents a substantial advance in the QM data infrastructure for biomolecular modeling. By delivering a chemically diverse, high‑accuracy, and consistently computed dataset that bridges the gap between small‑molecule QM benchmarks and protein‑fragment databases, it paves the way for next‑generation machine‑learning force fields capable of simulating complex biological systems with quantum‑level fidelity and classical‑scale efficiency.