Learning to Make Chemical Predictions: the Interplay of Feature Representation, Data, and Machine Learning Methods

Area of Science:

• * Cross-disciplinary applications in chemical, biological, and materials sciences.
• * Focus on the integration of machine learning with chemical insights.

Background:

• * Supervised machine learning (ML) is increasingly vital for scientific prediction.
• * Molecular property prediction is a key area benefiting from ML advancements.

Purpose of the Study:

• * To explore the synergy between ML methods and chemical descriptors.
• * To define data set requirements for accurate molecular property prediction.
• * To guide the selection of appropriate ML algorithms.

Main Methods:

• * Analysis of feature extraction and real-space representations for chemical structures.
• * Consideration of data set size and origin (experimental vs. computational).
• * Case study using Nuclear Magnetic Resonance (NMR) chemical shift prediction.

Main Results:

• * The choice of chemical descriptors significantly impacts prediction accuracy.
• * Data set characteristics (size, abundance, derivation) are critical for ML model performance.
• * Specific ML methods (deep learning, random forests, kernel methods) are suited to different data scenarios.

Conclusions:

• * Effective molecular property prediction requires careful consideration of descriptors, data, and ML algorithms.
• * NMR chemical shift prediction serves as a model for optimizing ML applications in chemistry.
• * Tailoring ML approaches to data availability and chemical representation is key to success.