DeeplyTough

DeeplyTough encodes three-dimensional protein binding sites with a convolutional neural network and compares resulting descriptor vectors via Euclidean distances to identify similar pockets for hit-finding, polypharmacology analysis, and protein function characterization.

Key Features:

• Convolutional Neural Network Architecture: Employs a convolutional neural network (CNN) to encode three-dimensional representations of protein binding sites into descriptor vectors and enables alignment-free comparison via pairwise Euclidean distances.
• Data-Driven Approach: Adopts a data-driven strategy leveraging large-scale benchmark datasets to address heterogeneity in traditional pocket-matching methods.
• Training Objectives: Trains descriptors so that similar binding pockets yield similar vectors, enforces a minimum margin between descriptors of dissimilar pockets, and optimizes robustness to nuisance variations.
• Benchmark Evaluation: Evaluated using three large-scale benchmark datasets, with strong performance on held-out data and competitive generalization to independently constructed datasets.

Scientific Applications:

• Hit-finding in drug discovery: Identifying similar binding pockets to suggest potential targets and guide lead discovery.
• Polypharmacology analysis: Analyzing how single drugs may interact with multiple protein targets by comparing pocket similarity.
• Protein function characterization: Assisting characterization of protein functions through structural comparison of binding sites.

Methodology:

Protein pockets are encoded into descriptor vectors using a CNN and compared alignment-free by computing pairwise Euclidean distances; training optimizes descriptor similarity for similar pockets, enforces margins for dissimilar pockets, and promotes robustness to nuisance variations.