(595j) Stability Prediction of Hypervalent Compaounds Based on Data-Centric Modelling

Hypervalent iodine compounds have become widely used reagents for the transfer of electro­philic substituents to donors. However, being very reactive by definition, many of these compounds are stable only because of a high barrier protecting them from transformation to chemically inactive forms.

In this study we explore hundreds of reagents of the general type shown in Figure 1 in view of their stability, i.e. the possibility of their use for the transfer of the substituent marked L.

For that matter, a support vector machine (SVM) was trained using simple descriptors based on the molecular structure and the atomic charges observed in the reactive center. The SVM, a popular machine learning tool for yes/no-type classification, one trained properly, was shown to successfully predict (in)stability of these compounds based on the descriptors selected.

In this talk we will focus on the selection of descriptors, the training of the SVM and its application to a large array of know and (still) unknown compounds. Relative to the explicit determination of the kinetic and thermodynamic stability, this approach allows a prediction at a small fraction of the cost since only the molecular structures and the atomic charges need to be available.

We will also show that in machine learning the availability of “negative” information, i.e. information on non-existing compounds, is essential. The fact that negative information is much more difficult to find, presents a challenge and calls for rethinking of the existing, success-based publication culture.

Figure 1: