Table 2.

Results on the RNAStrAlign dataset (sequence-wise CV)

	Precision	Recall	F1	Prec (S)	Recall (S)	F1 (S)	Weighted F1
Mfold [6]	0.450	0.398	0.420	0.463	0.409	0.433	0.366
RNAfold [8]	0.516	0.568	0.540	0.533	0.587	0.558	0.444
RNAstructure [7]	0.537	0.568	0.550	0.559	0.592	0.573	0.471
LinearFold [24]	0.620	0.606	0.609	0.635	0.622	0.624	0.509
CDPfold [12]	0.633	0.597	0.614	0.720	0.677	0.697	0.691
CONTRAfold [11]	0.608	0.663	0.633	0.624	0.681	0.650	0.542
E2Efold [13]	0.866	0.788	0.821	0.880	0.798	0.833	0.720
MXfold2 [16]	0.864	0.873	0.868	0.876	0.884	0.879	0.694
CNNFold + Argmax	0.955	0.861	0.900	0.955	0.872	0.902	0.812
CNNFold-mix + Argmax	0.956	0.912	0.932	0.958	0.915	0.934	0.863
CNNFold-mix + Blossom	0.975	0.907	0.936	0.978	0.909	0.938	0.872

Bold values are the best result in each column

“(S)” indicates the results when one-position shifts are allowed, that is for a base pair (i, j), the following predictions are also considered correct: $(i+1,j)$, $(i-1,j)$, $(i,j+1)$, $(i,j-1)$. The numbers for the comparison methods are from [13]. All trainable models have been trained on RSA-tr