The architecture of DeltaSplice. DeltaSplice comprises a feature-extraction module and multiple prediction modules. The feature-extraction module is constructed using residual-connected convolutional neural networks (CNNs), which convert a one-hot-encoded input sequence to a feature representation. Each prediction module consists of fully connected layers and a Softmax output layer that takes a feature representation as input and generates predictions for SSU or splice-site probabilities. The single-sequence mode employs two prediction modules to predict the SSU ${\hat{\mathit{u}}}_{\mathit{s}}$ and the splice-site probabilities ${\hat{\mathit{p}}}_{\mathit{s}}$ for each site in the input gene sequence s, based on the corresponding feature representation v_s. In the dual-sequence mode, the feature-extraction module calculates the feature representation ${\mathit{v}}_{{\mathit{s}}_{\mathrm{t}}}$ and ${\mathit{v}}_{{\mathit{s}}_{\mathrm{r}}}$ separately for the target gene sequence s_t and the reference gene sequence s_r. The predicted SSU ${\hat{\mathit{u}}}_{{\mathit{s}}_{\mathrm{t}}}$ for every site in the target gene sequence is computed using a prediction module, from the input ${\mathit{v}}_{{\mathit{s}}_{\mathrm{t}}}-{\mathit{v}}_{{\mathit{s}}_{\mathrm{r}}}+{\mathit{v}}_{{\mathit{u}}_{\mathrm{r}}}.$ Here ${\mathit{v}}_{{\mathit{u}}_{\mathrm{r}}}$ is the feature representation of the reference SSU u_r. RNA-seq data from adult brain tissues of humans and seven other mammalian species, as summarized in Supplemental Table S1, were used to estimate SSU values for model training.