RCL model. (A) The raw input is processed to extract α-length input segments representing the same genomic region in all replicates. (B) The α-length input segments are fed to encoder e( · ) to compute the cross-replicate contrastive loss. Then the embedding is fed to a multilayer perceptron (MLP), specifically a fully connected neural network, for class similarity loss and a decoder for the autoencoder (MSE) loss. The encoder/decoder has five ResNET blocks. (C) Shaded red boxes represent the elements contrasted in the respective losses.