Table 4. Summary of the independent benchmark RB44 results in terms of MCC.

Evaluation	Method	Accuracy	Specificity [+] (Precision)	Sensitivity (Recall)	F-measure	MCC	AUC
	aaRNA	0.823	0.551	0.643	0.593	0.483	0.845
	BindN+	0.835	0.614	0.468	0.531	0.439	0.819
Residue-	RNAbindR 2.0	0.805	0.514	0.532	0.523	0.401	0.801
based	Seq-CTRL	0.804	0.510	0.600	0.552	0.430	0.807
	KYG	0.771	0.449	0.638	0.527	0.392	0.808
	DRNA	0.788	0.480	0.660	0.556	0.430	N/A
	OPRA	0.746	0.403	0.551	0.465	0.311	N/A
	aaRNA	0.793	0.477	0.625	0.525	0.395	0.819
	BindN+	0.755	0.429	0.699	0.520	0.380	0.791 *
Protein-	RNAbindR 2.0	0.737	0.415	0.593	0.474	0.326	0.761 **
based	Seq-CTRL	0.763	0.459	0.547	0.473	0.343	0.782 ***
	KYG	0.727	0.397	0.672	0.486	0.334	0.775 ****
	DRNA	0.776	0.482	0.618	0.521	0.400	N/A
	OPRA	0.727	0.346	0.467	0.362	0.211	N/A

The same RNA-binding residue distance cutoff of 3.5 Å was used. Two evaluation methods (residue-based and protein-based) are used to estimate the performance of different predictors. Because the output of DRNA and OPRA methods provides no score describing residues’ RNA-binding propensities, an ROC analysis cannot be performed to estimate their AUCs. Except the DRNA method evaluated on a protein basis, which got a slightly higher MCC, aaRNA achieved better MCCs and AUCs than other sequence or structure-based methods, both in residue-based and protein-based performance evaluation. Paired Wilcoxon tests on protein-averaged AUCs of aaRNA and other methods indicated significant differences (P* < 3e-4, P** < 8e-7, P*** < 5e-8 and P**** < 2e-4).