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

The ability to predict log D directly from spectral patterns marks a conceptual shift in cheminformatics. In this work, we demonstrate that ¹H and ¹³C NMR spectra, computationally generated from molecular structures and transformed into machine learning-compatible vectors, can approach and rival classical structure-based descriptors such as ECFP4 fingerprints in modeling the log D parameter. Through comprehensive benchmarking of nearly 70 models across seven algorithmic classes and three pH conditions, we show that concatenation of ¹H and ¹³C NMR spectra offers the best trade-off between accuracy and efficiency. In the best case, a fused spectral CNN model achieved a root-mean-square error (RMSE) of 0.57 and a Q² of 0.76 using a 400-dimensional input vector─closely matching the ECFP4 benchmark (RMSE 0.56, Q² 0.78) despite being five times smaller. These findings challenge the assumption that descriptor richness must come at the cost of dimensional complexity. SHAP-based analysis revealed modality-specific patterns: ¹³C regions linked to aromatic and carbonyl carbons (110-170 ppm) increased predicted log D, while ¹H signals associated with polar groups, including OH, NH, amides, and ethers (2-4.5 and ∼8 ppm), reduced it. This positions NMR-based vectors as both interpretable and scalable alternatives to conventional fingerprints. By releasing a standalone graphical prediction tool based on our models, we make this paradigm practically accessible for real-world applications. This study establishes in silico-generated NMR spectra as valid and powerful descriptors in predictive modeling, paving the way for spectrum-driven approaches to drug discovery and property prediction.