Table 2.

Performance evaluation of m⁶A site prediction methods

		Performance on independent dataset (AUC)
Model	Method	A549	CD8T	Hela	HEK293 (sysy)	HEK293 (abacm)	MOLM13	Average AUC
Full Transcript	WHISTLE	0.965	0.930	0.953	0.936	0.968	0.933	0.948
	MethyRNA*	0.807	0.800	0.741	0.848	0.778	0.765	0.790
	SRAMP	0.856	0.841	0.762	0.883	0.838	0.759	0.761#
Mature mRNA	WHISTLE	0.903	0.904	0.894	0.936	0.818	0.823	0.880
	MethyRNA	0.751	0.734	0.676	0.848	0.698	0.686	0.732
	SRAMP	0.814	0.796	0.702	0.869	0.796	0.710	0.706#

Note: *The MethyRNA approach uses sequence-derived features with SVM (24), which we reproduced faithfully with the same training data of WHISTLE for comparison.

^#The SRAMP method was originally trained on A549, CD8T, HEK293 (sysy) and HEK293 (abacm). To avoid overfitting, only Hela and MOLM13 were considered when evaluating its average performance.

Only the m⁶A sites not previously used as training data were considered during performance evaluation, so the training sites and testing sites have no overlap. Please see Supplementary Table S4 for the results when all sites from the independent testing samples were considered.