Rapid and Accurate Prediction of pK_a Values of C–H Acids Using Graph Convolutional Neural Networks

Abstract

The ability to estimate the acidity of C−H groups within organic molecules in non-aqueous solvents is important in synthetic planning to correctly predict which protons will be abstracted in reactions such as alkylations, Michael additions, or aldol condensations. This Article describes the use of the socalled graph convolutional neural networks (GCNNs) to perform such predictions on the time scales of milliseconds and with accuracy comparing favorably with state-of-the-art solutions, including commercial ones. The crux of the method is to train GCNNs using descriptors that reflect not only topological but also chemical properties of atomic environments. The model is validated against adversarial controls, supplemented by the discussion of realistic synthetic problems (on which it correctly predicts the most acidic protons in >90% of cases), and accompanied by a Web application intended to aid the community in everyday synthetic planning.

Holistic Prediction of the pKa in Diverse Solvents Based on a Machine‐Learning Approach

“…However, all these ML models are only applicable for aqueous p K a s and ML models for non‐aqueous p K a prediction with reasonable accuracy remain virtually underdeveloped. Very recently, Grzybowski and co‐workers developed a prediction model for p K a in DMSO with MAE 2.1 p K a units by using graph convolutional neural networks (GCNNs), the model was trained with a small data set (817 p K a s) composed with half experimental data and half DFT (Density Functional Theory) calculated ones and is limited to only C−H acidity prediction in DMSO …”

Holistic Prediction of the pKa in Diverse Solvents Based on a Machine‐Learning Approach

“…Very recently,Grzybowski and co-workers developed aprediction model for pK a in DMSO with MAE 2.1 pK a units by using graph convolutional neural networks (GCNNs), the model was trained with as mall data set (817 pK a s) composed with half experimental data and half DFT (Density Functional Theory) calculated ones and is limited to only C À Ha cidity prediction in DMSO. [10] In the past decade,w eh ave engaged in collecting and curating accurate bond energies.I n2 016, we established au ser-friendly internet-based databank of pK a and BDE data: iBonD, [3] which is freely available at http://ibond. nankai.edu.cn.…”

Co‐crystal Prediction by Artificial Neural Networks**

“…Whereas the initial preprocessing step of the molecular fingerprint-based model type (FP;F igure 1a)c onsists of traditional hidden layers,t he molecular graph-based model type uses graph convolutional layers [50][51][52][53][54] (GCN;Figure 1b). Due to their wide applicability and strong performance, [55][56][57][58][59][60][61][62][63][64][65] graph convolutional neural networks have gained much popularity in recent years,a s they can learn specific molecular fragments and variations thereof that are decisive for the prediction of ap roperty, [50] rather than relying on statically defined atom combinations such as in circular fingerprints.Asimilar methodology could be used for the prediction of solvates,a nd it could be extended to applications involving two or more atomic or molecular species (e.g.m etal-organic frameworks).…”