FIGURE 1.

Architecture of network components in our study. Each colored bar represents a layer of operation in the network as indicated by the enclosed description. The numbers on the top or the side of bars specify the number of filters or hidden units in the corresponding layers. Activation functions following convolutional or fully connected layers and that in GRU are not shown. The former uses rectified linear function, while in the latter, hyperbolic tangent function is employed. ${[d_1\cdots d_{50}]}^{50}$ represents a numerical vector of size of 50 and c (v ₁, v ₂) denotes the concatenation of two vectors: v ₁ and v ₂. y _i ’s represent methylation state of the single CpG to predict in multiple profiles. When transferring trained models from source to target, components of Sequence, Methylation, Joint, or their combinations are transferred. Components that are not transferred are trained from scratch (i.e., random initialization). Conv, convolution; Max Pool, max pooling; Batch Norm, batch normalization; FC, fully connected.