Table 1.

Method performances on TIS-containing data.

Parametrization	Method	TP	FP	FN	Sn	Sp
	1st-ATG	18,553	1,150	0	0.9416	0.9416
	TISHunter	17,789	1,914	0	0.9029	0.9029
	ATGpr	17,160	2,543	0	0.8709	0.8709
	TIS Miner	15,521	3,650	532	0.7877	0.8096
	NetStart	5,123	14,527	53	0.2600	0.2607
homogeneous	LLKR	9,268	9,318	1,117	0.4704	0.4987
	WLLKR	12,511	4,486	2,706	0.6350	0.7361
	MFCWLLKR	15,167	4,535	1	0.7698	0.7698
	PFCWLLKR	14,692	4,191	820	0.7457	0.7781
	BAYES	10,121	6,482	3,100	0.5137	0.6096
cluster-specific	LLKR	11,964	6,946	793	0.6072	0.6327
	WLLKR	14,931	3,085	1,687	0.7578	0.8288
	MFCWLLKR	16,576	3,127	0	0.8413	0.8413
	PFCWLLKR	16,209	2,834	660	0.8227	0.8512
	BAYES	12,399	4,988	2,316	0.6293	0.7131
random split	LLKR	9,191	9,402	1,110	0.4665	0.4943
	WLLKR	12,491	4,507	2,705	0.6340	0.7349
	MFCWLLKR	15,183	4,519	1	0.7706	0.7706
	PFCWLLKR	14,648	4,198	857	0.7434	0.7772
	BAYES	10,084	6,509	3,110	0.5118	0.6077

19,703 TIS-containing instances were used in three separate five-fold cross-validation experiments. Results are shown from applying a non-stratified parameter set (homogeneous), a priori-known cluster-specific parameter sets for k = 3 (cluster-specific), and group-specific parameter sets for a random three-way split of the data (random split). TP represents the number of instances for which the method correctly identified a TIS; FP for which a prediction was made, though incorrect; and FN for which no prediction was made, but should have been (see Figure 2). $Sn=\frac{TP}{TP+FP+FN}$, and $Sp=\frac{TP}{TP+FP}$.