AnnoPRO: a strategy for protein function annotation based on multi-scale protein representation and a hybrid deep learning of dual-path encoding

Table 6 A comparison among those performances of AnnoPRO and two state-of-the-art methods (DeepGOPlus and PFmulDL) on constructing annotation models based on the benchmark named ‘PROBE’ in the original study [32], which consisted of 20,421 unique human proteins of distinct sequences

Method/Tool	BP		CC		MF
Method/Tool	F_max	AUPRC	F_max	AUPRC	F_max	AUPRC
DeepGOPlus	0.584	0.574	0.645	0.712	0.683	0.687
PFmulDL	0.533	0.526	0.623	0.682	0.648	0.651
AnnoPRO	0.643	0.664	0.652	0.717	0.709	0.709

By following the same criterion (using Oct 22, 2019 as a cutoff date) as that used by CAFA4 for data partitioning, 18,058 proteins were adopted as ‘Training and Validation’ data for model construction and 2,363 proteins were used as ‘Independent Testing’ dataset. The AnnoPRO, DeepGOPlus, and PFmulDL models were then retrained using these partitioned data. The values indicating the best performance among three methods were highlighted in BOLD, and AnnoPRO performed the best in all GO classes (BP, CC, MF) under both evaluating criteria (F_max, AUPRC). BP biological process, CC cellular component, MF molecular function

ISSN: 1474-760X