ICAN

ICAN identifies drug-target protein interactions (DTIs) from drug SMILES and protein amino acid sequences using an interpretable cross-attention neural network to improve prediction accuracy and mechanistic interpretability.

Key Features:

• Chemo-genomics feature-based modeling: Uses chemo-genomics feature-based methods that combine chemical and genomic descriptors for DTI prediction.
• Input descriptors: Trains models on drug representations from SMILES and target representations from amino acid sequences.
• Machine and deep learning models: Employs machine learning and deep learning architectures for DTI prediction.
• Attention-based mechanism: Implements attention mechanisms to enhance predictive performance and provide interpretability.
• Attention architecture exploration: Compares cross-attention versus self-attention, varies attention layer depths, and evaluates different context matrix selections.
• Optimal attention configuration: Identifies a simple attention strategy that decodes drug-related protein context features without protein-related drug context features as superior.
• Interpretability via attention weights: Statistically links weighted sites in the cross-attention weight matrix to experimental binding sites.
• Benchmarking on DAVIS: Demonstrates performance that outperforms state-of-the-art methods across multiple metrics on the DAVIS dataset.

Scientific Applications:

• DTI identification for drug discovery: Predicts drug-target interactions to support drug discovery and drug repositioning efforts.
• Mechanistic insight and binding-site mapping: Provides interpretable attention weights that correspond to experimental binding sites to elucidate interaction mechanisms.
• Benchmarking and model evaluation: Serves as a comparative method for performance evaluation on benchmark datasets such as DAVIS.

Methodology:

Models are trained using chemo-genomics descriptors from SMILES and amino acid sequences and employ attention-based neural network architectures with experiments comparing cross-attention versus self-attention, varying attention depths, and different context matrix selections, identifying a drug→protein decoding attention configuration as optimal.