DeepDTAF

DeepDTAF predicts protein–ligand binding affinity using a deep learning architecture that integrates protein-binding-pocket local inputs and global sequence-level contextual features to model multiscale interactions.

Key Features:

• Integration of Local and Global Contextual Features: Combines protein-binding-pocket local inputs with global sequence-level contextual features to enhance affinity prediction.
• Utilization of Sequence-Level Features: Operates on sequence-level data, enabling prediction when 3D protein structures are unavailable.
• Incorporation of Dilated Convolution: Uses dilated convolution to capture multiscale long-range interactions within protein–ligand complexes.
• Improved Prediction Accuracy: Demonstrated improved binding-affinity prediction accuracy compared to state-of-the-art methods.

Scientific Applications:

• Drug Discovery and Lead Identification: Predicts protein–ligand binding affinities to prioritize and identify promising drug candidates.
• Analysis with Limited Structural Data: Enables affinity prediction in research contexts where 3D structural information is scarce by relying on sequence-level features.

Methodology:

DeepDTAF implements a deep learning model that integrates local protein-binding-pocket features and global sequence-level contextual information using dilated convolution to capture multiscale interactions; the model was evaluated by comparative benchmarking against state-of-the-art methods and by component-level analysis.