Journal: Journal of Chemical Information and Modeling

Article Title: From NMR to AI: Do We Need 1 H NMR Experimental Spectra to Obtain High-Quality logD Prediction Models?

doi: 10.1021/acs.jcim.4c02145

Figure Lengend Snippet: Boxplots of absolute errors between computer-generated and experimental predictions of CHI logD (top) and Chrom logD (bottom) at different pH. In each plot, the first box corresponds to experimental 1 H NMR spectra, whereas the remaining ones are the methods used to obtain the generated 1 H NMR spectra. In each box, the horizontal dashed line describes the mean value, while the solid line refers to the median. Charts with marked outliers can be found in the Supporting Information .

Article Snippet: The change in RMSE error values for a set of 60% contribution of predicted spectra relative to experimental 1 H NMR spectra to a set with 100% contribution of theoretical spectra for CHI logD models averaged −15.2% and −11.0% for Chrom logD.

Techniques: Generated

Journal: Journal of Chemical Information and Modeling

Article Title: From NMR to AI: Do We Need 1 H NMR Experimental Spectra to Obtain High-Quality logD Prediction Models?

doi: 10.1021/acs.jcim.4c02145

Figure Lengend Snippet: Distribution of the absolute prediction error for Gradient Boosting models as a function of the predicted parameter. The upper panel corresponds to the CHI logD parameter at three different pH values: 2.6, 7.4, and 10.5; whereas the lower panel represents the Chrom logD parameter at the same pH values. The data series indicates the source of input data ( 1 H NMR spectra) used for training the machine learning models.

Article Snippet: The change in RMSE error values for a set of 60% contribution of predicted spectra relative to experimental 1 H NMR spectra to a set with 100% contribution of theoretical spectra for CHI logD models averaged −15.2% and −11.0% for Chrom logD.

Techniques:

Journal: Journal of Chemical Information and Modeling

Article Title: From NMR to AI: Do We Need 1 H NMR Experimental Spectra to Obtain High-Quality logD Prediction Models?

doi: 10.1021/acs.jcim.4c02145

Figure Lengend Snippet: Panel presents the mean RMSE values obtained for the Gradient Boosting models using the 10CV method for mixed input datasets. Each graph illustrates the trend of mean RMSE values as the contribution of generated 1 H NMR spectra of a given type increases within the set of experimental 1 H NMR spectra. The upper panel displays data for the CHI logD parameter, while the lower panel corresponds to the Chrom logD parameter. The series within each panel represents the modeled parameter values at different pH levels.

Article Snippet: The change in RMSE error values for a set of 60% contribution of predicted spectra relative to experimental 1 H NMR spectra to a set with 100% contribution of theoretical spectra for CHI logD models averaged −15.2% and −11.0% for Chrom logD.

Techniques: Generated

Journal: Journal of Chemical Information and Modeling

Article Title: From NMR to AI: Do We Need 1 H NMR Experimental Spectra to Obtain High-Quality logD Prediction Models?

doi: 10.1021/acs.jcim.4c02145

Figure Lengend Snippet: Bar charts present CHI logD parameter RMSE metrics for the validation query of the trained Gradient Boosting models for a given type of 1 H NMR input data. The query was performed using two sets of 1 H NMR spectra corresponding to structural motifs that were absent in the training and testing datasets. The last bar in each chart shows data obtained from logD value predictions using Instant JChem, as the CHI logD value can be directly compared with logD.

Article Snippet: The change in RMSE error values for a set of 60% contribution of predicted spectra relative to experimental 1 H NMR spectra to a set with 100% contribution of theoretical spectra for CHI logD models averaged −15.2% and −11.0% for Chrom logD.

Techniques: Biomarker Discovery

Journal: Journal of Chemical Information and Modeling

Article Title: From NMR to AI: Do We Need 1 H NMR Experimental Spectra to Obtain High-Quality logD Prediction Models?

doi: 10.1021/acs.jcim.4c02145

Figure Lengend Snippet: Bar plots illustrate the performance metrics (RMSE and NRMSE) for Gradient Boosting models trained to predict CHI logD and Chrom logD at three different pH levels (2.6, 7.4, and 10.5). Each pair of bars corresponds to CHI logD (plain bars) and Chrom logD (hatched bars) for each of the four input data sources: Experimental 1 H NMR spectra were obtained using three spectral generation methods (DFT, JASON, and NMRshiftDB2). The metrics are displayed in four categories: RMSE, NRMSE scaled by the mean of the target values ( y̅ ), NRMSE scaled by the range of the target values ( y max – y min ), and NRMSE scaled by the standard deviation of the target values (σ).

Article Snippet: The change in RMSE error values for a set of 60% contribution of predicted spectra relative to experimental 1 H NMR spectra to a set with 100% contribution of theoretical spectra for CHI logD models averaged −15.2% and −11.0% for Chrom logD.

Techniques: Standard Deviation

Journal: Journal of Chemical Information and Modeling

Article Title: From NMR to AI: Do We Need 1 H NMR Experimental Spectra to Obtain High-Quality logD Prediction Models?

doi: 10.1021/acs.jcim.4c02145

Figure Lengend Snippet: Statistical Metrics of Datasets for the CHI and Chrom LogD at Different pH Levels

Article Snippet: The change in RMSE error values for a set of 60% contribution of predicted spectra relative to experimental 1 H NMR spectra to a set with 100% contribution of theoretical spectra for CHI logD models averaged −15.2% and −11.0% for Chrom logD.

Techniques: