LTPConstraint: a transfer learning based end-to-end method for RNA secondary structure prediction

BMC Bioinformatics

Table 8 Contrast experiments with other models

Length	Family	LTPConstraint			CONTRAfold			LinearFold			ProbKnot			RNAfold			CycleFold
Length	Family	Precision	Recall	F1-score	Precision	Recall	F1-score	Precision	Recall	F1-score	Precision	Recall	F1-score	Precision	Recall	F1-score	Precision	Recall	F1-score
Encoding_128	Rfam	0.8599	0.7897	0.8233	0.5016	0.6186	0.5540	0.5429	0.5502	0.5465	0.4705	0.6118	0.5319	0.4598	0.5965	0.5193	0.2887	0.5329	0.3745
	5SrRNA	0.9857	0.9804	0.9831	0.6541	0.7337	0.6916	0.7330	0.7306	0.7318	0.5748	0.6036	0.5888	0.5716	0.6401	0.6039	0.3094	0.4984	0.3818
	tRNA	0.9985	0.9992	0.9988	0.7047	0.7787	0.7399	0.7445	0.7504	0.7474	0.6777	0.7691	0.7205	0.6659	0.7378	0.7000	0.3097	0.4695	0.3732
	PDB	0.6695	0.3050	0.4190	0.0180	0.0102	0.0130	0.0142	0.0071	0.0095	0.0228	0.0128	0.0164	0.0175	0.0105	0.0132	0.0324	0.0274	0.0297
	SPR	0.9929	0.9971	0.9950	0.6730	0.7496	0.7092	0.6990	0.6763	0.6875	0.6338	0.7319	0.6793	0.6355	0.7208	0.6755	0.3182	0.4865	0.3848
Encoding_512	grpl	0.8304	0.8894	0.8589	0.6589	0.6509	0.6549	0.6713	0.5557	0.6080	0.6124	0.6401	0.6259	0.6013	0.6383	0.6192	0.2068	0.1055	0.1397
	RNP	0.5334	0.7000	0.6054	0.5952	0.5998	0.5975	0.5335	0.4654	0.4972	0.5548	0.5383	0.5464	0.4956	0.5813	0.5350	0.2931	0.2008	0.2383
	SRP	0.7130	0.7378	0.7252	0.5731	0.6279	0.5992	0.6204	0.6099	0.6151	0.5693	0.6179	0.5926	0.5670	0.6234	0.5938	0.2621	0.3810	0.3105
	telomerase	0.3752	0.8728	0.5248	0.4327	0.6083	0.5057	0.4334	0.5670	0.4913	0.4026	0.5483	0.4643	0.3912	0.5572	0.4597	0.1193	0.2330	0.1578
	tmRNA	0.7550	0.8767	0.8113	0.4367	0.4592	0.4477	0.4208	0.3618	0.3891	0.3862	0.4350	0.4092	0.3828	0.4378	0.4085	0.1549	0.2107	0.1786

Six RNA secondary structure prediction models including LTPConstraint were tested using the processed validation set described in “Data collection and processing” section. Since LTPConstraint uses transfer learning, we need to use the pre-trained model. For the 5 families of Encoding_128 (see Table 4), LTPConstraint uses pre-trained model trained by using Rfam_128 data, while for the 5 families of Encoding_512, LTPConstraint uses the pre-trained model trained from Rfam_512 data

ISSN: 1471-2105