Skip to main content
Fig. 4 | Journal of Cheminformatics

Fig. 4

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

Fig. 4

Explaining a misclassification of a N-sulfoglucosamine sulfohydrolase (EC 3.10.1) as a glycosidase (EC 3.2.1). The model ECXRhea correctly predicts the reaction (1) to be catalysed by a hydrolase (EC 3), primarily focusing on the water (OH2) and the hydrolysed bond, both with a positive contribution towards EC 3. In addition, there is a small negative contribution against EC3 shown on the amine group. Unlike 1, where positive and negative contributions are shown in one reaction drawing, positive and negative contributions are split into separate depictions for 2 and 3 for visualization purposes. For the top prediction (67.1%) of model ECXYZRhea the focus of the model shifts to a non-reactive site including a hydroxy group in the N-sulfo-D-glucosamine as a major positive contribution (2a), while the sulfur-nitrogen bond is the major negative contribution (2b). For the correct prediction (top-2, 12.8%), the model remains focused on the hydrolised sulfur-nitrogen bond with a positive contribution (3a) as the negative contributions (3b) can be found on the D-glucosamine and the proton

Back to article page