Ontology-aware deep learning enables ultrafast and interpretable source tracking among sub-million microbial community samples from hundreds of niches

Table 1 Evaluation of ONN4MST on all five datasets

Dataset	No. biomes	No. samples	All features					Selected features
Dataset	No. biomes	No. samples	Pr	Rc	Acc	F_max	AUC	Pr	Rc	Acc	F_max	AUC
Combined	114	125,823	0.826	0.662	0.995	0.740	0.971	0.868	0.774	0.997	0.820	0.977
Human	25	53,553	0.822	0.521	0.984	0.695	0.972	0.894	0.826	0.991	0.863	0.984
Water	44	27,667	0.842	0.766	0.992	0.803	0.966	0.854	0.764	0.992	0.813	0.971
Soil	16	11,528	0.915	0.778	0.986	0.850	0.974	0.892	0.881	0.989	0.890	0.982
FEAST	3	10,270	0.793	0.795	0.984	0.803	0.980	0.895	0.812	0.989	0.862	0.991

ONN4MST achieved the accuracy higher than 0.98 for all five datasets, and the AUC higher than 0.97 for all five datasets. For each dataset, we used the model trained on that dataset for evaluation. The evaluation procedure of the ONN model is described in the “Methods” section. ONN4MST based on all features and selected features were both evaluated at the bottom (sixth) layer with a threshold of 0.5
Abbreviations: Pr precision, Rc recall, Acc accuracy

ISSN: 1756-994X