An explainability framework for deep learning on chemical reactions exemplified by enzyme-catalysed reaction classification

Table 1 Accuracies and F-scores of the different models trained on Rhea and ECREACT

Model	Accuracy	F-Score	Training Time	Energy Use
ECX_Rhea	\(0.98\pm eq0.00\)	\(0.97\pm 0.01\)	100 s	43 Wh
ECXY_Rhea	\(0.96\pm 0.01\)	\(0.88\pm 0.02\)	160 s	70 Wh
ECXYZ_Rhea	\(0.95\pm 0.00\)	\(0.87\pm 0.02\)	190 s	82 Wh
ECX_ECREACT	\(0.98\pm 0.00\)	\(0.96\pm 0.00\)	2090 s	904 Wh
ECXY_ECREACT	\(0.95\pm 0.00\)	\(0.82\pm 0.01\)	2170 s	936 Wh
ECXYZ_ECREACT	\(0.93\pm 0.00\)	\(0.77\pm 0.01\)	2810 s	1,216 Wh

The energy use is calculated based on the energy use of the device (Dell XPS 15, i7-12700 H CPU, NVIDIA GeForce RTX 3050 Ti Laptop GPU) and includes the power usage of models trained for 4x cross-validation and four experiments with fingerprint variations. The hyperparameter values were taken from the previous work on organic reactions [44]. The total resulting energy consumption for model experimentation, training, and validation for this project was 3.25 kWh. Energy mix (2022): 65% hydro, 23% solar, and 12% other renewables

ISSN: 1758-2946