Table 2 Micro- and macro-averaged F1 scores for all models across EC levels 1-4 on the test subset restricted to sequences with less than 50% sequence identity and less than 80% coverage relative to the training set

From: Interpretable Kolmogorov-Arnold networks for enzyme commission number prediction

Metric

Level

CLEAN

DeepECtransformer

DeepEC

  

MLP

KAN

MLP

KAN

MLP

KAN

Micro F1-scores

1

0.818 ± 0.015

0.821 ± 0.021*

0.511 ± 0.070

0.661 ± 0.061*

0.601 ± 0.025

0.717 ± 0.019*

2

0.727 ± 0.019

0.731 ± 0.029*

0.486 ± 0.071

0.641 ± 0.068*

0.576 ± 0.024

0.711 ± 0.020*

3

0.685 ± 0.014

0.701 ± 0.019*

0.476 ± 0.072

0.629 ± 0.070*

0.553 ± 0.019

0.698 ± 0.012*

4

0.589 ± 0.009

0.592 ± 0.015*

0.454 ± 0.071

0.601 ± 0.065*

0.521 ± 0.014

0.673 ± 0.013*

Macro F1-scores

1

0.785 ± 0.011

0.798 ± 0.019*

0.513 ± 0.069

0.663 ± 0.062*

0.591 ± 0.022

0.718 ± 0.017*

2

0.595 ± 0.008

0.612 ± 0.015*

0.309 ± 0.053

0.499 ± 0.050*

0.372 ± 0.020

0.521 ± 0.012*

3

0.429 ± 0.007

0.453 ± 0.012*

0.206 ± 0.034

0.349 ± 0.031*

0.257 ± 0.027

0.368 ± 0.021*

4

0.237 ± 0.007

0.249 ± 0.009*

0.086 ± 0.015

0.167 ± 0.014*

0.120 ± 0.016

0.183 ± 0.012*

  1. Asterisks (*) denote statistically significant improvements by the Wilcoxon signed-rank test; boldface indicates the best performing variant (MLP or KAN) within each model.