Fig. 1

AUC of various alternative splicing models. (a) A convolutional DNN trained on sequences up to 384 bp from each of the four splice sites involved in cassette splicing (8 × 384 bp). (b) Same as (a), weighting the post-convolutional feature map by conservation (8 × 384 bp). (c) Same as (b), using only the 100 bp at each end of the cassette exon (2 × 100 bp). (d) Same as (c), using only the first 100 bp of each flanking intron (2 × 100 bp). (e) State of the art: the method of [11] (1393 features). (f) A DNN trained on only the conservation of the regions in (d) (2 × 100 features). (g) A DNN trained on junction conservation divided by average conservation over 100 bp (2 features). (h) A DNN trained on junction conservation (2 features). (i) A DNN trained on average conservation over 100 bp (2 features). (j) A DNN trained on the combined features of (e) and (g) (1393 + 2 features)