Exploring the structural basis of organic compounds by predicting experimental IR peaks: a machine learning analysis

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Materials Science

Background:

• Understanding the structure-property relationships of organic compounds is vital for advancements in chemistry and materials science.
• Infrared (IR) spectroscopy is a key technique for identifying functional groups and elucidating molecular structure.
• Predicting experimental spectral data computationally can accelerate research and discovery.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting experimental carbonyl peaks in the IR spectrum of organic compounds.
• To identify key molecular descriptors that correlate with IR spectral features.
• To assess the synthetic accessibility of predicted compounds.

Main Methods:

• Utilized Modred and RDKit molecular descriptors as input features for machine learning models.
• Employed Extra Trees (ET) regression for predicting IR spectral peaks.
• Applied SHapley Additive exPlanations (SHAP) to interpret model predictions and identify influential descriptors.
• Performed hyperparameter tuning to optimize model performance, specifically focusing on the number of estimators.

Main Results:

• FractionCSP3 emerged as the most correlating descriptor for both Modred and RDKit descriptor sets.
• Extra Trees regression demonstrated strong predictive performance, achieving a coefficient of determination (R²) between 0.72 and 0.78.
• SHAP analysis identified BCUT2D_MRLOW (RDKit) and FCSP3 (Modred) as the most influential descriptors.
• Optimized models with 50 estimators and calculated synthetic accessibility (SA) scores within a 0.00-0.15 range.

Conclusions:

• Machine learning models, particularly Extra Trees regression, can effectively predict experimental IR carbonyl peaks in organic compounds.
• Molecular descriptors like FractionCSP3, BCUT2D_MRLOW, and FCSP3 are crucial for understanding the structural basis of IR spectral properties.
• The integration of synthetic accessibility scores provides valuable insights into the practical feasibility of synthesizing predicted compounds.