TL;DR
This paper tackles the challenge of generating accurate molecular conformations by comparing experimental crystal structures with computationally derived gas-phase structures. They found that quantum chemical calculations (GFN2-xTB) can reliably predict torsional preferences, offering a powerful alternative to relying solely on limited crystallographic data, and paving the way for new, data-driven conformer sampling methods.
What Problem Does This Paper Address?
Generating accurate 3D molecular conformations is crucial for many areas of chemistry, from drug discovery to materials science. A common approach is to use experimental crystallographic data to derive "torsional preferences" – the preferred angles around chemical bonds. However, this method has limitations: it's restricted to molecules that are stable, synthetically accessible, and can be crystallized, leading to biases and gaps in the available data. This means that less common or more complex molecules might not be well-represented. The authors aim to determine if computationally generated gas-phase conformers, obtained using modern semiempirical quantum chemical methods, can provide a more comprehensive and unbiased source of torsional preferences, overcoming the limitations of experimental data.
Methods
The core of this study involves extensive computational sampling of molecular conformations. The researchers used the GFN2-xTB method, a fast and accurate semiempirical quantum chemical approach, to perform exhaustive torsion sampling on over 3 million organic small molecules. For each molecule, they generated conformers and identified the lowest-energy one. They then compared the torsional preferences derived from these gas-phase computed conformers with those obtained from experimental crystallographic data. To validate their findings, they also performed density functional theory (DFT) calculations (@B97X-D/def2-SVP) on a subset of molecules to confirm the accuracy of the GFN2-xTB method. Finally, they explored fitting torsional distributions using Gaussian functions, proposing this as a more robust approach than traditional Fourier-based methods for developing new conformer sampling tools.
Datasets
The study utilized several large datasets:
- Crystallographic Open Database (COD): Approximately 88,000 organic compounds from experimental crystal structures.
- PubChemQC: Over 3 million molecules with pre-computed gas-phase theoretical geometries from density functional theory.
- Pitt Quantum Repository: Approximately 88,000 molecules.
- ZINC: A subset of about 70,000 molecules.
- Platinum ligand database: Around 4,000 molecules.
The authors note that the PubChemQC set compounds tend to be smaller than those in other datasets. They also mention that their work is comparable to other large-scale computational efforts like ANI, GEOM, QCArchive, QMugs, and SPICE.
Key Insights and Ideas
- GFN2-xTB is a reliable source for torsional preferences: The study demonstrates that GFN2-xTB calculations can accurately capture torsional preferences, providing a valuable alternative to experimental data. This is a significant finding because it opens up the possibility of exploring a much wider chemical space.
- Experimental data has inherent biases: The paper highlights how limitations in crystallization and synthetic accessibility restrict the types of molecules found in crystallographic databases, leading to potential biases in derived torsional preferences.
- Gas-phase vs. solid-state differences: Differences observed between gas-phase and experimental data are often attributed to intermolecular interactions and packing effects in the solid state, which are absent in isolated gas-phase calculations. This is particularly evident when comparing to bound-ligand geometries.
- Gaussian fits outperform traditional methods: Fitting torsional distributions with Gaussian peaks proved to be more accurate than traditional cosine Fourier analysis (used in ETKDG), especially for capturing subpopulations of conformers. This suggests a more nuanced way to represent torsional preferences.
- Towards data-driven conformer generation: The work lays the groundwork for developing new conformer sampling methods, like the proposed "quantum torsion-derived distance geometry" (QTDG), that are less reliant on experimental data and can handle a broader range of chemical structures.
Key Results
- Geometric RMSD: CREST/GFN2 ensembles showed high fidelity compared to experimental geometries, performing better than the widely used ETKDG method for molecules with few rotatable bonds on the COD dataset.
- Radius of Gyration: A notable difference was observed in the radius of gyration between Platinum set data and CREST/GFN2 conformers, especially for molecules with more rotatable bonds, suggesting a preference for extended conformations in bound ligands due to intermolecular interactions.
- Torsional Preference Correlation: A good correlation (median r² of 0.61 for acyclic and 0.83 for ring torsions) was found between experimental COD torsional preferences and computed gas-phase preferences. Poor correlations were often linked to a lack of sufficient experimental data.
- DFT Validation: DFT calculations confirmed that GFN2-xTB torsion angles are highly accurate, with mean absolute deviations typically below 5°, suggesting that GFN2 is sufficient for generating reliable torsional distributions.
- Fitting Accuracy: Gaussian fits achieved median r² values above 0.9 for both acyclic and ring torsional patterns, significantly outperforming ETKDG and cosine fits.
Limitations and Future Work
- Computational Cost: While GFN2-xTB is fast, it still requires a median of 1 hour per compound for widespread conformational sampling, which can be a limitation.
- Intermolecular Interactions: The study acknowledges that gas-phase calculations do not account for intermolecular interactions, which can influence conformational preferences in condensed phases or when bound to biological targets. Future work should consider benchmarks that capture these effects, such as rotational spectroscopy.
- Correlated Torsions: Sampling ring torsion angles in isolation is challenging due to their correlation with other torsions in the ring. Future work should focus on refining methods that handle these correlations, potentially through distance geometry approaches.
- QTDG Development: The authors propose the development of a Quantum Torsion-Derived Distance Geometry (QTDG) method as an alternative to ETKDG, which would leverage quantum calculations and avoid reliance on experimental data. This is a key direction for future research.
- Expanding Datasets: The paper suggests that future work could involve comparisons with geometric information from gas-phase spectroscopy and further refinement of torsion patterns using additional conformer sampling for sparse datasets.
Summary updated