Reaction Classification and Yield Prediction using the Differential Reaction Fingerprint DRFP

Probst, Daniel; Schwaller, Philippe; Reymond, Jean‐Louis

doi:10.33774/chemrxiv-2021-mc870

Cited by 11 publications

(18 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It is noteworthy that except for the descriptors about the reactants, the dipole moments of the solvents also remarkably affect the reaction yield. Reaxtica achieved better performance (R 2 = 0.87) than the DL-based method rxnfp 26 or fingerprint-based DRFP 27 using the HTE dataset. We further divided the USPTO and Reaxys dataset into 8:1:1 for training, validation and test, respectively.…”

Section: Yield Of Suzuki-miyaura Coupling Reactionmentioning

confidence: 94%

“…23 Regarding to SMILES-based models, Schwaller et al have applied natural language processing technology in reaction outcomes predictions, 24 which also extended to the selectivity of carbohydrate reactions 25 and reaction yield. 26,27 These DL-based models worked well in forward reaction outcome prediction in USPTO and HTE datasets. However, most of them are still black-box models that do not generalize well.…”

Section: Introductionmentioning

confidence: 91%

See 1 more Smart Citation

Reaxtica: A knowledge-guided machine learning platform for fast and accurate reaction selectivity and yield prediction

Lin

et al. 2022

Preprint

View full text Add to dashboard Cite

Reaction selectivity and yield prediction are important for chemical synthesis. Most existing computational methods use either computational expensive and complicated quantum mechanics-based models that are not easy for experimental chemists to use or black-box deep learning models that do not generalize well outside of the training space and lack explanation. Herein, using convenient physics-based electronic descriptors and structure-based steric descriptors, we developed an explainable machine learning platform, Reaxtica, that outperformed previous methods in four different reaction types and tasks, including regioselectivity, site-selectivity, enantioselectivity, and yield predictions. Further descriptor analysis helps understand reaction mechanisms behind the data. As a practical and robust toolbox, Reaxtica can be easily applied to different chemical reactions and extended to out-of-sample reaction. To assist chemists’ daily research, we further built an easy-to-use webserver, which only takes seconds to run and can be accessed at http://www.pkumdl.cn:8000/reaxtica/.

show abstract

Section: Yield Of Suzuki-miyaura Coupling Reactionmentioning

confidence: 94%

Section: Introductionmentioning

confidence: 91%

Reaxtica: A knowledge-guided machine learning platform for fast and accurate reaction selectivity and yield prediction

Lin

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…[15] The reactants and product fingerprint consist of a 512-bit extended connectivity fingerprints with a radius of 3, [16] concatenated to a 512-bit RDKit fingerprints with a maximum path length of 7. [17] Other featurization schemes of chemical reactions have been suggested recently, [18,19] but their use is outside the scope of our study.…”

Section: Methodsmentioning

confidence: 99%

Prediction of the Chemical Context for Buchwald‐Hartwig Coupling Reactions

Genheden

Mårdh

Lahti

et al. 2022

Molecular Informatics

View full text Add to dashboard Cite

We present machine learning models for predicting the chemical context for Buchwald-Hartwig coupling reactions, i. e., what chemicals to add to the reactants to give a productive reaction. Using reaction data from inhouse electronic lab notebooks, we train two models: one based on single-label data and one based on multi-label data. Both models show excellent top-3 accuracy of approximately 90 %, which suggests strong predictivity. Furthermore, there seems to be an advantage of including multi-label data because the multi-label model shows higher accuracy and better sensitivity for the individual contexts than the single-label model. Although the models are performant, we also show that such models need to be re-trained periodically as there is a strong temporal characteristic to the usage of different contexts. Therefore, a model trained on historical data will decrease in usefulness with time as newer and better contexts emerge and replace older ones. We hypothesize that such significant transitions in the context-usage will likely affect any model predicting chemical contexts trained on historical data. Consequently, training context prediction models warrants careful planning of what data is used for training and how often the model needs to be re-trained.

show abstract

“…Modeling the yield of these datasets (4K C-N couplings 15 , or 2K Suzuki-Miyaura couplings in flow 14 ) produces predictive models with R 2 or AUROC > 0.9. 11,15,[27][28][29][30][31][32][33][34] However, models trained on these datasets demonstrate limited ability to extrapolate beyond the molecules in their training sets, in part due to the minimal structural diversity in the dataset.…”

Section: Introductionmentioning

confidence: 99%

Roadmap to Pharmaceutically Relevant Reactivity Models Leveraging High-Throughput Experimentation

Kalyani

Struble

et al. 2022

Preprint

View full text Add to dashboard Cite

The merger of High-Throughput Experimentation (HTE) and data science presents an opportunity to both accelerate and inspire innovations in synthetic chemistry. Similarly, developments in machine learning (ML) have enabled the distillation of large and complex data sets into predictive models capable of generalizing patterns in the data. However, efforts to merge HTE with ML remain constrained by a few reported datasets with limited structural diversity and corresponding trained models that do not extrapolate well to substrates beyond the training set. Herein, we detail the first ML models for Pd-catalyzed C–N couplings using pharmaceutically relevant structurally diverse large data sets (~ 5000 unique products) generated using nanomole scale compatible chemistry. Careful consideration is given to both the diversity of the data set and accurate model predictions for substrates bearing features beyond those present in the training set. The structural diversity in the data set is enabled by leveraging the Merck & Co., Inc Building Block Collection with an initial focus on C–N coupling using secondary amines. The large dataset enables the systematic evaluation of model performance using five different data-splitting strategies. These five splits are carefully designed to evaluate the model’s ability to extrapolate beyond the substrates in the training set. The accuracy of classification models built with a lens toward application to medicinal chemistry campaigns exceeded the baseline precision-recall by 25-67% depending on the splitting strategy. These results would manifest as significant enrichment of successful C–N couplings using the hits recommended by the models. In addition, the accuracy of the best models for each of the five splits ranges between 70-87% suggesting excellent overall predictivity of the models even for completely unseen substrates.

show abstract

Reaction Classification and Yield Prediction using the Differential Reaction Fingerprint DRFP

Cited by 11 publications

References 19 publications

Reaxtica: A knowledge-guided machine learning platform for fast and accurate reaction selectivity and yield prediction

Reaxtica: A knowledge-guided machine learning platform for fast and accurate reaction selectivity and yield prediction

Prediction of the Chemical Context for Buchwald‐Hartwig Coupling Reactions

Roadmap to Pharmaceutically Relevant Reactivity Models Leveraging High-Throughput Experimentation

Contact Info

Product

Resources

About