The testing data set was extracted as the set of all new genes from dbTSSv5 which is based on hg17 and did not appear in dbTSSv4 (training & validation data). 
Genes that have more than a 30% mRNA overlap are removed from consideration. 

The complete list of the 1024 testing genes were taken from (http://www.fml.tuebingen.mpg.de/raetsch/projects/arts)
We downloaded the corresponded sequences from (http://genome.ucsc.edu).

for every gene on the positive strand was downloaded as:
Min( start, tss) -2000 ----> end+1000

While from the negative strand the reverse complement was downloaded as:
Max(end, tss)+2000 ----> start -1000.

Note:
Training and validation dataset was downloaded from (http://www.fml.tuebingen.mpg.de/raetsch/projects/arts).

More details are available in the published paper.