Fig. 2

Constructing the embedded DeepPASS model for position-insensitive prediction of PASs. a Construction of known PAS set and stringent PAS set. Previously reported PASs in at least two databases of PolyA_DB3, PolyA-seq, and PolyASite (v2.0) or in the manually examined GENCODE annotation were combined to achieve known PAS set. Stringent PAS set was further constructed from known PASs that are annotated in all three databases or in the GENCODE annotation and was split with a 9:1 ratio between a model training set and an independent validation set for DeepPASS and DeepPAS-fixed models. b Schematic of DeepPASS construction and evaluation. Left, data processing strategy and model architecture. Middle, a sequence shifting strategy around stringent PASs was applied to construct positive training set for establishing DeepPASS model. Right, the generally used strategy with fixed sequences around stringent PASs for DeepPAS-fixed model. See “Methods” section for details. c The ROC curves of DeepPASS and DeepPAS-fixed to indicate their training performance. AUC values of DeepPASS (red) and DeepPAS-fixed (blue) were shown in plot. d The ROC curves of DeepPASS, DeepPASTA, and APARENT on the validation set. AUC values of DeepPASS (red), DeepPASTA (green), and APARENT (purple) were shown in plot from five independent validation repeats with very low standard errors. ***p < 0.001, statistical significance was assessed by Student’s t test. See “Methods” section for details. e Position-insensitive prediction of PASs with DeepPASS model. To assess positional tolerance of different models, 200-bp sequences shifting around the PASs in validation set with 5 bp stride were used to test accuracy of each model. The percentage represents true positive rate of each condition. DeepPASS (red) is more tolerant than DeepPAS-fixed (blue) and previously reported DeepPASTA (green) and APARENT (purple) models