From NMR to AI: Fusing ¹H and ¹³C Representations for Enhanced QSPR Modeling.

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Predicting physicochemical properties, such as log D (distribution coefficient), is crucial for drug discovery and development.
• Traditional methods rely on structure-based descriptors (e.g., ECFP4 fingerprints), which can be computationally intensive and lack interpretability.
• Nuclear Magnetic Resonance (NMR) spectra contain rich information about molecular structure and electronic environment.

Purpose of the Study:

• To investigate the potential of using computationally generated NMR spectra as descriptors for predicting log D.
• To compare the performance of NMR spectral descriptors against classical structure-based fingerprints.
• To develop and validate machine learning models for log D prediction using NMR spectral data.

Main Methods:

• Generated 1H and 13C NMR spectra computationally from molecular structures.
• Transformed NMR spectra into machine learning-compatible vectors.
• Benchmarked nearly 70 machine learning models across seven algorithmic classes and three pH conditions.
• Utilized Convolutional Neural Networks (CNNs) for spectral data analysis.
• Performed SHAP-based analysis for model interpretability.

Main Results:

• NMR spectral vectors achieved performance comparable to ECFP4 fingerprints in predicting log D.
• A fused spectral CNN model demonstrated strong predictive accuracy (RMSE 0.57, Q² 0.76) with a significantly smaller input vector size.
• Concatenation of 1H and 13C NMR spectra provided the best balance of accuracy and efficiency.
• SHAP analysis revealed specific NMR spectral regions correlating with log D values, enhancing model interpretability.

Conclusions:

• In silico-generated NMR spectra are effective and scalable descriptors for predictive modeling.
• NMR-based vectors offer a viable, interpretable, and efficient alternative to conventional fingerprints.
• This study paves the way for spectrum-driven approaches in drug discovery and property prediction.