Molecular Design Based on Integer Programming and Splitting Data Sets by Hyperplanes.

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Chemical informatics

Background:

• A novel framework exists for designing chemical compounds with desired properties using mixed integer linear programming (MILP).
• This framework relies on feature and prediction functions, with machine learning methods constructing the latter.
• Improving the learning performance of these prediction functions is crucial for the framework's efficacy.

Purpose of the Study:

• To enhance the learning performance of prediction functions within a molecular design framework.
• To develop a data splitting method to improve prediction accuracy for chemical properties.
• To facilitate the design of chemical compounds with specific, targeted properties.

Main Methods:

• A novel data splitting technique is proposed, dividing a dataset C into two subsets (C(i), i=1,2) using a hyperplane in chemical space.
• Compounds are partitioned based on observed values relative to a threshold θ, with subsets containing lower and higher values, respectively.
• A combined prediction function ψ is constructed by integrating individual prediction functions (ψi, i=1,2) trained on each subset.

Main Results:

• The proposed data splitting method demonstrated improved learning performance for prediction functions.
• Computational experiments showed enhanced accuracy in predicting several chemical properties.
• The method proved effective where previous prediction functions struggled to achieve good performance.

Conclusions:

• The developed data splitting strategy significantly enhances the performance of prediction functions in molecular design frameworks.
• This approach offers a viable solution for improving the prediction of challenging chemical properties.
• The findings contribute to more effective computational design of novel chemical compounds.