Sigma profiles in deep learning: towards a universal molecular descriptor.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Molecular descriptors are crucial for quantitative structure-activity relationships (QSAR) and property prediction.
• Sigma profiles offer a novel 3D representation of molecular electronic properties.

Purpose of the Study:

• To demonstrate the efficacy of sigma profiles as molecular descriptors in deep learning models.
• To predict a diverse range of physicochemical properties using sigma profiles.
• To extend model capabilities by including temperature as a predictive feature.

Main Methods:

• Utilized sigma profiles from 1432 compounds as input features.
• Trained convolutional neural networks (CNNs) to correlate and predict physicochemical properties.
• Adapted developed CNN architectures to incorporate temperature data.

Main Results:

• Achieved accurate correlations and predictions of multiple physicochemical properties.
• Demonstrated the effectiveness of sigma profiles in deep learning for chemical property prediction.
• Successfully integrated temperature as an additional predictive feature.

Conclusions:

• Sigma profiles are highly effective molecular descriptors for deep learning applications.
• The developed deep learning models accurately predict chemical properties and can be extended with environmental factors like temperature.