Fig. 1

Overview of the xCAPT5 Model Architecture, which encompasses five distinct phases denoted by the capital letters in the parenthesis. A Input Stage: The model takes two protein sequences as input. B Embedding Phase: The ProtT5-XL-UniRef50 Protein Language Model processes the sequences to produce amino acid embeddings. C Single Sequence Learning Phase: Subsequent to embedding, each sequence traverses through five convolutional modules. Within each module, four layers are executed in sequence: the first performs convolutions with kernel sizes 2 (conv 2), 3 (conv 3, not illustrated in the figure), and 4 (conv 4), generating varying feature maps. These maps are then activated via the Swish function in the second layer. The third layer acts on the activation output, applying average pooling (AP) and max pooling (MP) to retain the most important features. The fourth layer (Pooling accumulation by depth), functioning as an auxiliary pathway, applies global max pooling and global average pooling on activation output across different depths, followed by a multi-kernel concatenation (Multi-kernel concat) to create a comprehensive feature profile for each sequence. The concatenated outputs are processed through a two-layer feed-forward network incorporating fully connected layers (dense), ReLU activation, and drop out. D Sequence Pair Learning Phase: The extracted representations from individual sequences are combined and fed into a three-layer feed-forward network to learn the refined features of protein pairs. E Intermediate Phase: The XGBoost algorithm is employed to train on these integrated features, optimizing the model’s predictive capability. F Prediction Phase: The final output is a probabilistic score given by the trained XGBoost model, which predicts the interaction potential between the two input protein sequences