Predicting Reaction Yields via Supervised Learning

Zuranski, A.; Alvarado, Jesus I. Martinez; Shields, Benjamin J.; Doyle, Abigail G.

doi:10.1021/acs.accounts.0c00770

“…This puts forward higher requirements for the generalization ability of machine learning model. Żurański and co-workers proposed a cross-validation method called leave one molecule out, which helped to evaluate the ability of machine learning models to explore unknown chemical spaces 9 . Here we adopted the test method of leaving out a molecule, the training set did not contain the reactions of the molecule, and all the reactions of the molecule were put into the test set, to evaluated the extrapolation ability of ML models training in known chemical reaction space.…”

Section: Resultsmentioning

confidence: 99%

“…Prior leading efforts of reaction performance or reaction enantioselectivity prediction using machine learning had focused on partial molecular chemical space mainly represented by density functional theory (DFT) 3,5 and molecular ngerprint or SMILES based descriptors 7,8 . Wherein the DFT based descriptors offered more subtle physiochemical property differences compared to molecular ngerprint/SMILES ones, impelling it a dominant strategy 9 . Another pivotal challenge for the implementation of machine learning models to reaction prediction was its requirement of su cient and consistent experimental data 10 .…”

Section: Introductionmentioning

confidence: 99%

Highly accurate prediction of reaction performance and enantioselectivity with a uniform machine learning protocol

Sun

¹

,

Zhu

²

,

You

³

et al. 2021

Preprint

0

View full text Add to dashboard Cite

Synthetic reactions, especially asymmetric reactions are key components of modern chemistry. Chemists have put enormous experimental effort into recognizing various molecule patterns to enable efficient synthesis and asymmetric catalysis. Recent application of machine learning algorithms and chemoinformatics in this field demonstrated their huge potential in facilitating this process by accurate prediction. However, existing methods are relatively limited to specific designed data set, and only implement single prediction of reaction performance or reaction enantioselectivity, rendering their general use in broader scenarios challenging. Here we provide a uniform machine learning protocol that can predict both reaction performance and enantioselectivity with high accuracy. Reconstruction of molecular chemical space derived from more comprehensive three-dimensional atomic and molecular descriptors allow for training of our neural network-based model over four representative datasets. This uniform machine learning protocol was validated with outperformance of accuracy than other methods over all four cases (C-C, C-N, C-S cross coupling reactions and asymmetric hydrogenation) in the prediction of both reaction performance and enantioselectivity. It was also successfully applied to the out-of-set and sparse set prediction, leveraging its possible wide application in accelerating synthesis improve and molecular architects.

show abstract

“…At the fundamental level, reactions proceed through 3D interactions, and we therefore expect that explicitly modeling the 3D shape of reaction components could lead to better performance for reactivity tasks. These include but are not limited to reaction yield [37,51], selectivity [49,15], and condition prediction [28], and even retrosynthesis planning [38].…”

Section: Discussion and Outlookmentioning

confidence: 99%

3D Computer Vision Models Predict DFT-Level HOMO-LUMO Gap Energies from Force-Field-Optimized Geometries

Maser

¹

,

Reisman

²

2021

Preprint

0

View full text Add to dashboard Cite

We investigate 3D deep learning methods for predicting quantum mechanical energies at high-theory-level accuracy from inexpensive, rapidly computed molecular geometries. Using space-filled volumetric representations (voxels), we explore the effects of radial decay from atom centers and rotational data augmentation on learnability. We test several published computer vision models for 3D shape learning, and construct our own architecture based on 3D inception networks with physically meaningful kernels. We provide a framework for further studies and propose a modeling challenge for the computer vision and molecular machine learning communities.

show abstract

“…At the fundamental level, reactions proceed through 3D interactions, and we therefore expect that explicitly modeling the 3D shape of reaction components could lead to better performance for reactivity tasks. These include but are not limited to reaction yield [37,51], selectivity [49,15], and condition prediction [28], and even retrosynthesis planning [38].…”

Section: Discussion and Outlookmentioning

confidence: 99%

3D Computer Vision Models Predict DFT-Level HOMO-LUMO Gap Energies from Force-Field-Optimized Geometries

Maser

¹

,

Reisman

²

2021

Preprint

0

View full text Add to dashboard Cite

We investigate 3D deep learning methods for predicting quantum mechanical energies at high-theory-level accuracy from inexpensive, rapidly computed molecular geometries. Using space-filled volumetric representations (voxels), we explore the effects of radial decay from atom centers and rotational data augmentation on learnability. We test several published computer vision models for 3D shape learning, and construct our own architecture based on 3D inception networks with physically meaningful kernels. We provide a framework for further studies and propose a modeling challenge for the computer vision and molecular machine learning communities.

show abstract

Predicting Reaction Yields via Supervised Learning

Abstract: Technical analysis of the previously published random forest model and its ability to predict reactivity for unseen molecules.

Cited by 115 publications

References 58 publications

Highly accurate prediction of reaction performance and enantioselectivity with a uniform machine learning protocol

Highly accurate prediction of reaction performance and enantioselectivity with a uniform machine learning protocol

3D Computer Vision Models Predict DFT-Level HOMO-LUMO Gap Energies from Force-Field-Optimized Geometries

3D Computer Vision Models Predict DFT-Level HOMO-LUMO Gap Energies from Force-Field-Optimized Geometries

Contact Info

Product

Resources

About