Inverse analysis of QSPR using mol-infer

In J-OCTA’s MI-Suite, inverse analysis can be performed to predict molecular structures from physical properties. It is equipped with mol-infer, developed by the Haraguchi Laboratory at Kyoto University. First, an artificial neural network (ANN) is used to predict physical properties from molecular structures, and then mixed-integer linear programming (MILP) is applied for inverse computation. This enables estimation of molecular structures that match target properties, and structural isomers can also be obtained.

Use Cases Highlights

Prediction of molecular structure from physical property values
Inverse analysis using chemical graphs
Application to molecular design targeting partition coefficients as desired properties

Prediction of molecular structure from physical property values

Data for a partition coefficient (logP = 10.0) set as the target property for inverse analysis using mol-infer are shown. First, data for 1,297 molecular structures and property values are used to train an artificial neural network (ANN) on the structure–property relationship. Inverse analysis is then performed using the trained model.

Target property of mol-infer (partition coefficient = 10.0)

Inverse analysis using chemical graphs

The seed structures of molecular structures used for inverse analysis and tree structures corresponding to functional groups are shown. By combining these structural elements using mixed-integer linear programming (MILP), molecular structures that satisfy target properties are searched.

Seed structure (left) and tree structure (right)

Application to molecular design targeting partition coefficients as a property

An example of molecular structures obtained by inverse analysis is shown. Structural isomers are also included, and forward property estimation was performed again for the obtained structures. The partition coefficient was 9.8, close to the target value, confirming the effectiveness of the inverse analysis.