Multi-task learning for pKa prediction

Abstract

Many compound properties depend directly on the dissociation constants of its acidic and basic groups. Significant effort has been invested in computational models to predict these constants. For linear regression models, compounds are often divided into chemically motivated classes, with a separate model for each class. However, sometimes too few measurements are available for a class to build a reasonable model, e.g., when investigating a new compound series. If data for related classes are available, we show that multi-task learning can be used to improve predictions by utilizing data from these other classes. We investigate performance of linear Gaussian process regression models (single task, pooling, and multitask models) in the low sample size regime, using a published data set (n = 698, mostly monoprotic, in aqueous solution) divided beforehand into 15 classes. A multi-task regression model using the intrinsic model of co-regionalization and incomplete Cholesky decomposition performed best in 85 % of all experiments. The presented approach can be applied to estimate other molecular properties where few measurements are available.

Empirical Conversion of pKa Values between Different Solvents and Interpretation of the Parameters: Application to Water, Acetonitrile, Dimethyl Sulfoxide, and Methanol

“…To convert a p K a value between two solvents, one needs a single additive shift parameter for each molecular family and pair of solvents. Such families of titratable molecular groups are previously used in the context of empirical prediction of p K a values in aqueous solvent, , and family-specific linear functions serve for interconversion between the estimated p K a values.…”

“…Subsequently, the measured and ETM p K a values are used to establish the ECM for the five molecular families. The same operation is performed employing the measured p K a values in DMSO and the p K a values in water obtained with the JPM . Here, the quality of the ECM is evaluated by comparing the ECM p K a values in water with the corresponding ETM and JPM p K a values.…”

Prediction of pKa Using Machine Learning Methods with Rooted Topological Torsion Fingerprints: Application to Aliphatic Amines

“…A wide variety of methods have been developed for p K a prediction − and they have been comprehensively reviewed. − One of the common approaches is the quantitative structure–property relationship (QSPR). QSPR methods are a class of empirical approaches.…”

“…RTorsion was introduced before in the design of a fluorinated fragment library by capturing the local structure environments around F/CF3. 53 It was also applied to predict the 19 F NMR chemical shift by a distance-weighted K-nearest neighbor (KNN) algorithm. 54 In this study, we used RTorsion in QSPR modeling for pK a .…”

Semisupervised Multitask Learning With Gaussian Processes

“…The collocated transfer approach that we employ was investigated on [5] and is based on the IMC approach of [3] and [4], which as discussed in Section III represents an intermediate level of transfer compared with other methods. Although the IMC model has been found useful in real applications such as estimating the molecular properties of compounds [33], or in other forms of transfer learning such as meta-generalizing [34], it should be made clear that we do not claim, NOR do we believe, that the IMC model is necessarily superior to other transfer learning approaches on every data set; the purpose of this paper was to present a methodology for inference in this important model class, and to illustrate and contrast its working on a number of examples and with a number of methods. Continuing, we presented two possible approaches: in the static geometry case, tasks are assumed to have independent geometric structures, so that transfer of geometry can only happen through the correlations between tasks.…”