Table 3. The performance of five predictive models.

The average AUPRC, precision, recall, AUROC, and the average training time of the five predictive models from 5-fold cross-validation are described. For the donor sites, the HSplice tool was used as a benchmark.

Site	Model	AUPRC		Precision		Recall		AUROC		Runtime(Colab)
		mean	SD	mean	SD	mean	SD	mean	SD
Donor	CNN_3	0.986	0.0005	0.936	0.0013	0.979	0.0009	0.989	0.0003	12 m
	CNN_4	0.986	0.0002	0.930	0.0015	0.982	0.0010	0.989	0.0001	12 m
	CNN_LSTM	0.983	0.0004	0.932	0.0003	0.975	0.0013	0.986	0.0002	25 m
	SVM	0.923	0.0007	0.937	0.0007	0.968	0.0012	0.952	0.0006	2 hra
	RF	0.913	0.0004	0.939	0.0006	0.942	0.0007	0.940	0.0002	11 s
	HSplice	0.968		0.928		0.936		0.975		N/A
Acceptor	CNN_3	0.979	0.0003	0.910	0.0028	0.968	0.0027	0.982	0.0004	12 m
	CNN_4	0.979	0.0008	0.905	0.0030	0.973	0.0012	0.982	0.0006	12 m
	CNN_LSTM	0.975	0.0008	0.914	0.0020	0.960	0.0013	0.979	0.0006	25 m
	SVM	0.893	0.0017	0.915	0.0018	0.948	0.0013	0.930	0.0013	2.30 hra
	RF	0.866	0.0009	0.910	0.0011	0.893	0.0020	0.902	0.0010	11 s

Notes.

^aThe SVM-based model was run on a local laptop without GPU (Intel Core i5-3230M CPU 2.60 GHz, x64-based Windows OS, 8 GB of RAM, 256 GB SSD).

Bold styling emphasizes the highest values regarding the evaluation metrics used in the study.