Skip to main content
. 2026 May 13;15:giag056. doi: 10.1093/gigascience/giag056

Figure 2.

A six-panel figure displaying performance evaluation charts for the EC-LMGraph model. The top row, comprising panels A through C, focuses on newly annotated enzyme structures. Panel A is a box plot showing precision, recall, and F1 scores categorized by the four EC class levels. Panel B is a similar box plot evaluating these same three metrics across five confidence thresholds ranging from 0.5 to 0.9. Panel C is a line graph comparing the F1 scores of four different models against maximum percentage sequence identity, with the EC-LMGraph line charting the highest performance. The bottom row, comprising panels D through F, displays grouped bar charts benchmarking twelve different methods on newly annotated enzyme sequences. Panel D groups the bars by precision, recall, and F1 score. Panel E groups the bars by the 1st, 2nd, 3rd, and 4th EC digits to compare F1 scores across the models. Panel F groups the bars by maximum percentage sequence identity to compare F1 scores. A shared legend at the bottom identifies the color coding for the twelve computational methods compared in the bottom row.

EC-LMGraph prediction performance. (A–C) Benchmarking results on newly annotated enzyme structures. (A) Performance of EC-LMGraph on the prediction of EC main class, subclass, sub-subclass, and sub-sub-subclass numbers, adopting a confidence threshold of 0.5. (B) Dependence of precision, recall, and F1 scores of EC-LMGraph on confidence thresholds ranging from 0.5 to 0.9. (C) Dependence of prediction performances on the training set maximum percentage of sequence identity cut-off for the compared methods. (D–F) Benchmarking results on newly annotated enzyme sequences. (D) and (E) Prediction performances on protein sequences with newly annotated EC numbers for the various compared methods. The input protein structures for the structure-based methods (i.e., EC-LMGraph, DeepFRI, and COFACTOR) were AlphaFold2-predicted structures. For (F), the dependence of prediction performances for 3rd/4th-digit EC numbers on the training set maximum percentage of sequence identity cut-off is shown for the compared methods.