DeepCDA

DeepCDA predicts compound–protein binding affinities across different data domains using deep learning and adversarial domain adaptation to improve cross-domain generalization for drug discovery.

Key Features:

• Cross-Domain Adaptation: Handles scenarios where training and test datasets originate from different distributions to improve generalization.
• Encoder Network: Uses a deep learning encoder trained with convolutional layers and long-short-term memory (LSTM) layers to learn representations of compounds and proteins.
• Two-Sided Attention Mechanism: Incorporates a two-sided attention mechanism to encode interaction strengths between protein and compound substructures.
• Adversarial Domain Adaptation: Employs an adversarial domain adaptation technique to learn a feature encoder network tailored for the test domain.
• Affinity Prediction: Applies the learned test encoder network to predict binding affinities for new compound–protein pairs.
• Evaluation Datasets: Validated using KIBA, Davis, and BindingDB datasets.

Scientific Applications:

• Drug Discovery: Predicts binding affinities to aid identification of potential therapeutic candidates.
• Cross-Domain Affinity Assessment: Enables reliable affinity prediction when training and test data have differing distributions.
• Compound–Protein Interaction Prioritization: Ranks promising compound–protein interactions for experimental follow-up.

Methodology:

Uses a three-phase computational pipeline: (1) train an encoder network with convolutional layers and LSTM layers to learn compound and protein representations and apply a two-sided attention mechanism, (2) perform adversarial domain adaptation to obtain a feature encoder for the test domain, and (3) apply the learned test encoder to predict binding affinities; evaluation was performed on KIBA, Davis, and BindingDB.