Table 1.

Summary statistics as percentages for sensitivity, specificity, precision, recall, ROC-AUC, and PR-AUC for the predicted annotation

Annotation level	Performance metric	1st quartile	Median	Mean	3rd quartile
Class	Sensitivity/recall	99.01	99.50	99.31	99.96
	Specificity	99.95	99.98	99.96	100.00
	Precision	77.73	88.33	87.56	100.00
	ROC-AUC	99.94	99.98	99.97	100.00
	PR-AUC	80.32	92.86	88.95	100.00
Group	Sensitivity/recall	99.76	100.00	98.91	100.00
	Specificity	99.93	100.00	99.92	100.00
	Precision	74.74	100.00	85.65	100.00
	ROC-AUC	99.98	100.00	99.92	100.00
	PR-AUC	78.30	100.00	88.90	100.00
Mechanism	Sensitivity/recall	99.72	100.00	97.50	100.00
	Specificity	99.94	99.98	99.95	100.00
	Precision	73.79	88.17	85.58	100.00
	ROC-AUC	99.98	100.00	99.81	100.00
	PR-AUC	76.39	94.57	87.21	100.00
Model	Sensitivity/recall	99.72	100.00	97.50	100.00
	Specificity	97.43	99.93	100.00	100.00
	Precision	49.98	100.00	75.73	100.00
	ROC-AUC	99.97	100.00	99.91	100.00
	PR-AUC	52.78	100.00	79.80	100.00

All metrics excluding the AUC were calculated at the E-value threshold of 1e-25, which optimized the PR curve. We note that other thresholds might be useful depending on the false-positive tolerance of the use-case