Extending the Applicability of the ANI Deep Learning Molecular Potential to Sulfur and Halogens

Devereux, Christian; Smith, Justin S.; Davis, Kate K.; Barros, Kipton; Zubatyuk, Roman I.; Isayev, Olexandr; Roitberg, Adrián E.

doi:10.1021/acs.jctc.0c00121

Cited by 259 publications

(380 citation statements)

References 64 publications

Supporting

Mentioning

377

Contrasting

Order By: Relevance

“…[47,48] Energies are read from all output files using the cclib [49] version 1.6.2 and pybel version 3.0. [50] Machine learning methods included "bag-of-features" representations and ANI-1x [51] , ANI-1ccx [52] , and ANI-2x [53] models. The Bag-of-Features representations chosen were Bag of Bonds [54] (BOB), Bond Angle Torsion [55] (BAT), and Bond Angle Torsion Typed (BATTY).…”

Section: Methodsmentioning

confidence: 99%

See 1 more Smart Citation

Assessing conformer energies using electronic structure and machine learning methods

Folmsbee

Hutchison

2020

Int J of Quantum Chemistry

View full text Add to dashboard Cite

We have performed a large-scale evaluation of current computational methods, including conventional small-molecule force fields; semiempirical, density functional, ab initio electronic structure methods; and current machine learning (ML) techniques to evaluate relative single-point energies. Using up to 10 local minima geometries across 700 molecules, each optimized by B3LYP-D3BJ with single-point DLPNO-CCSD(T) triplezeta energies, we consider over 6500 single points to compare the correlation between different methods for both relative energies and ordered rankings of minima. We find that the current ML methods have potential and recommend methods at each tier of the accuracy-time tradeoff, particularly the recent GFN2 semiempirical method, the B97-3c density functional approximation, and RI-MP2 for accurate conformer energies. The ANI family of ML methods shows promise, particularly the ANI-1ccx variant trained in part on coupled-cluster energies. Multiple methods suggest continued improvements should be expected in both performance and accuracy.

show abstract

Section: Methodsmentioning

confidence: 99%

“…Machine learning methods included “bag‐of‐features” representations and ANI‐1x [ 51 ] , ANI‐1ccx [ 52 ] , and ANI‐2x [ 53 ] models. The Bag‐of‐Features representations chosen were Bag of Bonds [ 54 ] (BOB), Bond Angle Torsion [ 55 ] (BAT), and Bond Angle Torsion Typed (BATTY).…”

Section: Methodsmentioning

confidence: 99%

Assessing conformer energies using electronic structure and machine learning methods

Folmsbee

Hutchison

2020

Int J of Quantum Chemistry

View full text Add to dashboard Cite

show abstract

“…We selected metabolites containing only elements which could be subjected to energy calculation by ANI-2x potentials (CHONSFCl) to yield 7547 molecules. 10 The number of rotatable bonds were calculated for each metabolite to remove trivial or unfeasible cases for conformation generation. Metabolites with rotatable bonds on the range of the 50th to 95th quantile were retained, resulting in all metabolites having four to fourteen rotatable bonds.…”

Section: Metabolite Curationmentioning

confidence: 99%

“…Recent development and benchmarking of machine learning based potentials such as ANI-2x have prompted its utilization for certain high throughput quantum chemical applications. 10,11 For example, ANI-1ccx potentials were used to accelerate the refinement of generated conformers in a quantum mechanical (QM) NMR spectral prediction workflow. 12,13 The conformers optimized by these models, however, generally do not converge to the same local minima as ab initio methods.…”

Section: Introductionmentioning

confidence: 99%

AutoGraph: Autonomous Graph-Based Clustering of Small-Molecule Conformations

Tanemura

Das

Merz

2021

J. Chem. Inf. Model.

View full text Add to dashboard Cite

While accurately modeling the conformational ensemble is required for predicting properties of flexible molecules, the optimal method of obtaining the conformational ensemble seems as varied as their applications. Ensemble structures have been modeledby generation, refinement, and clustering of conformations with a sufficient number of samples. We present a conformational clustering algorithm intended to automate the conformational clustering step through the Louvain algorithm, which requires minimal hyperparameters and importantly no predefined number of clusters or threshold values. The conformational graphs produced by this method for O-succinyl-L-homoserine, oxidized nicotinamide adenine dinucleotide, and 200 representative metabolites each preserved the geometric/energetic correlation expected for points on the potential energy surface. Clustering based on these graphs provide partitions informed by the potential energy surface. Automating conformational clustering in a workflow with AutoGraph may mitigate human biases introduced by guess-and-check over hyperparameter selection while allowing flexibility to the result by not imposing predefined criteria other than optimizing the model's loss function.

show abstract

“…The training set for ANI-2x adds the additional elements of F, Cl, and S while providing additional torsion sampling data. 5 The BAND NN model uses a subset of the ANI-1 data set with only nonequilibrium geometries with energies within 30 kcal/mol of the equilibrium energy. Although these methods have been shown to perform adequately in their respective papers, the range for bond stretch applications has been limited to the harmonic portion of the potential energy curve, rarely examining the potential energy curves further from equilibrium.…”

Section: Introductionmentioning

confidence: 99%

Evaluation of Thermochemical Machine Learning for Potential Energy Curves and Geometry Optimization

2021

View full text Add to dashboard Cite

While many machine learning methods, particularly deep neural networks, have been trained for density functional and quantum chemical energies and properties, the vast majority of these methods focus on single-point energies. In principle, such ML methods, once trained, o↵er thermochemical accuracy on par with density functional and wave function methods but at speeds comparable to traditional force fields or approximate semiempirical methods. So far, most e↵orts have focused on optimized equilibrium single-point energies and properties. In this work, we evaluate the accuracy of several leading ML methods across a range of bond potential energy curves and torsional potentials. Methods were trained on the existing ANI-1 training set, calculated using the !B97X / 6-31G(d) single points at non-equilibrium geometries. We find that across a range of small molecules, several methods o↵er both qualitative accuracy (e.g., correct minima, both repulsive and attractive bond regions, anharmonic shape, and single minima) and quantitative accuracy in terms of the mean absolute percent error near the minima. At the moment, ANI-2x, FCHL, and a new libmolgrid-based convolutional neural net show good performance.

show abstract

Extending the Applicability of the ANI Deep Learning Molecular Potential to Sulfur and Halogens

Cited by 259 publications

References 64 publications

Assessing conformer energies using electronic structure and machine learning methods

Assessing conformer energies using electronic structure and machine learning methods

AutoGraph: Autonomous Graph-Based Clustering of Small-Molecule Conformations

Evaluation of Thermochemical Machine Learning for Potential Energy Curves and Geometry Optimization

Contact Info

Product

Resources

About