Limitations of current machine learning models in predicting enzymatic functions for uncharacterized proteins.
-
YegV (b2100) is a member of a sugar kinase superfamily where DeepECTF incorrectly predicted EC 2.7.1.92 (dehydro-2-deoxygluconokinase), an activity that actually belongs to KdgK (b3526)
"For example, b2100 was annotated as a dehydro-2-deoxygluconokinase (EC 2.7.1.92) using DeepECTF and as an uncharacterized sugar kinase YegV (EC 2.7.1.-) in UniProt (Supplementary Table 1b)"
-
The specific substrate of YegV is unknown; only the first three EC digits (2.7.1) are reliably predicted, while the fourth digit specifying substrate was incorrectly propagated from a paralog
"These 2 predictions differ by the fourth or last position of the EC number that specifies substrate specificity. This protein is a member of a superfamily of sugar kinases with multiple nonisofunctional paralogous subgroups that phosphorylate different substrates"
-
The dehydro-2-deoxygluconokinase activity is encoded by KdgK/b3526, not YegV, and the misprediction illustrates overpropagation error from nonisofunctional paralogs
"The dehydro-2-deoxygluconokinase (EC 2.7.1.92) activity is encoded by another member of this superfamily KdgK/b3526 (Supplementary Table 1d, bottom). Here, DeepECTF predicted correctly the first 3 digits of the EC number but not the last, making an overpropagation mistake (error 6 in Table 1 and Fig. 1)"