Skip to main content
. 2021 Oct 28;38(3):597–603. doi: 10.1093/bioinformatics/btab746

Fig. 2.

Fig. 2.

Masked language modeling. Positive and negative sites are indicated in green and blue, respectively. Sites to train on (orange) are either masked (80%) or randomized (20%). The model is optimized to infer the original methylation state given the corrupted input using the cross-entropy loss