Permutationally Invariant Deep Learning Approach to Molecular Fingerprinting with Application to Compound Mixtures.

Area of Science:

• Chemistry
• Machine Learning
• Deep Learning

Background:

• Deep learning algorithms are increasingly applied to chemical synthetic planning and reaction prediction.
• Supervised learning, a subset of deep learning, is used to predict chemical properties like reaction yields and types.
• Chemical fingerprints are commonly used as descriptors for reaction property prediction, but their permutational invariance is understudied.

Purpose of the Study:

• To investigate the impact of permutational invariance in chemical fingerprints for deep learning models.
• To improve the accuracy of reaction property predictions using deep learning.
• To develop a model capable of accurately predicting hydrogen peroxide loss in complex chemical formulations.

Main Methods:

• Utilizing chemical fingerprints with inherent permutational invariance.
• Applying deep learning architectures to these invariant fingerprints.
• Training and validating models on established datasets and a novel dataset for hydrogen peroxide loss prediction.

Main Results:

• Consistent improvements in prediction accuracy compared to previous studies were observed by incorporating permutational invariance.
• The developed model accurately predicted hydrogen peroxide loss using a dataset with over 20 ingredients per formulation.
• The study demonstrates the efficacy of permutational invariance in enhancing deep learning models for chemical predictions.

Conclusions:

• Permutational invariance is a crucial factor for improving the performance of deep learning models in chemical reaction prediction.
• The methodology presented offers a robust approach for predicting chemical properties, including complex degradation processes like hydrogen peroxide loss.
• This work highlights the potential of invariant deep learning representations for advancing chemical informatics and formulation science.