Prediction of Molar Extinction Coefficients using Directed Message-Passing Neural Networks : Augmenting Graph-Based Encodings with Semi-Empirical Quantum Chemistry Features

Sammanfattning

The determination of reaction yield is a crucial analysis step when synthesizing novel compounds. Because each compound requires dedicated calibration experiments to determine its molar extinction coefficient (MEC), this process is highly resource-intensive. Consequently, traditional yield determination becomes infeasible in high-throughput settings. In this study, a calibration-free workflow was proposed based on a multi-target Chemprop model trained to predict MEC and λmax directly from molecular structure data, with semi-empirical xTB/sTDA predictions incorporated as additional input features. The model was trained on the Deep4Chem dataset and tested on an independent set of 209 chromophores from the PhotochemCAD™ database. The Chemprop models demonstrated strong comparative performance for λmax on the independent test set, outperforming both the random forest baseline and a previously published model while reducing the parameter count by 97%. While the models successfully captured structural trends for λmax, the prediction error (RMSE ≈ 100 nm) remains too high for direct operational use. This limitation may be partly attributed to systematic inconsistencies in how absorption peaks are reported across datasets. Furthermore, predicting the MEC presented inherent modeling challenges. While the strictly negative R2 scores for MEC were exacerbated by the narrow variance of log-MEC values in the test set, the magnitude of the error confirms that current models cannot yet replace calibration experiments. Specifically, an RMSE of roughly 0.5 log units corresponds to a threefold error in MEC. Achieving usable generalization performance will require additional training data within the relevant chemical space. Furthermore, the current models are limited to predicting MEC at λmax, whereas high-throughput settings typically measure absorbance at a fixed wavelength. However, wavelength-specific MEC estimation is feasible by scaling semi-empirical xTB/sTDA spectra with the model-predicted MEC, offering a practical path toward calibration-free quantification in automated high-throughput workflows.