Machine learning-based approaches for ubiquitination site prediction in human proteins

BMC Bioinformatics

Table 6 Sequence-based results based on the macro-F1 score on the test set using a weighted loss strategy. The bold values show the best results

Window size	LSTM	BERT-small	BERT-tiny	Nystromformer	SqueezeBERT
5	0.532 ± 0.003	0.543 ± 0.013	0.541 ± 0.001	0.543 ± 0.001	0.545 ± 0.001
7	0.526 ± 0.002	0.524 ± 0.005	0.529 ± 0.002	0.527 ± 0.001	0.519 ± 0.007
9	0.53 ± 0.001	0.536 ± 0.001	0.534 ± 0.001	0.536 ± 0.003	0.533 ± 0.002
15	0.545 ± 0.001	0.55 ± 0.001	0.548 ± 0.002	0.549 ± 0.001	0.549 ± 0.002
21	0.555 ± 0.003	0.558 ± 0.002	0.555 ± 0.002	0.556 ± 0.003	0.556 ± 0.003
27	0.563 ± 0.003	0.562 ± 0.002	0.562 ± 0.003	0.56 ± 0.002	0.561 ± 0.002
33	0.563 ± 0.004	0.566 ± 0.004	0.564 ± 0.002	0.566 ± 0.003	0.563 ± 0.002
45	0.571 ± 0.002	0.566 ± 0.003	0.568 ± 0.002	0.563 ± 0.006	0.566 ± 0.001
55	0.572 ± 0.003	0.569 ± 0.005	0.57 ± 0.003	0.573 ± 0.005	0.565 ± 0.002
77	0.574 ± 0.004	0.572 ± 0.003	0.571 ± 0.003	0.567 ± 0.01	0.569 ± 0.004
99	0.573 ± 0.002	0.57 ± 0.004	0.596 ± 0.004	0.571 ± 0.007	0.57 ± 0.001
Avg (5–99)	0.555 ± 0.009	0.556 ± 0.017	0.558 ± 0.008	0.556 ± 0.015	0.554 ± 0.01
Avg (7–99)	0.557 ± 0.008	0.557 ± 0.011	0.56 ± 0.008	0.557 ± 0.015	0.555 ± 0.01

ISSN: 1471-2105