Model Definition

• Input: Two drug fingerprints, each of size 1024.
• Encoder: Two-layer fully connected neural network with ReLU activation.
• Classifier: Merges encoded fingerprints and predicts interaction severity class (3 classes).
• Data-split: Stratified split on 80% - train, 10% - validation, 10% - test
• Training Epochs: 100 unless reached convergence in validation accuracy
• Learning Rate: 0.001
• Batch Size: 2048