DFCNN

DFCNN predicts protein-ligand interactions using a Deep Feature Convolutional Neural Network to enable large-scale virtual screening in drug discovery.

Key Features:

• High efficiency: Achieves success rates ~22-fold higher than random chance at a score cutoff of 0.99, and in some cases up to ~98-fold higher than random guessing.
• Scalability and speed: Processes protein-compound predictions in ~0.0000225 seconds per prediction on 80 Intel CPU cores (2.00 GHz) with 60 GB RAM, achieving speeds tens of thousands of times faster than AutoDock Vina or Schrödinger high-throughput virtual screening.
• Dataset reduction: Reduces candidate sets by approximately 20- to 100-fold for downstream docking or molecular dynamics (MD) simulation steps.
• Experimental validation: Predictions have been experimentally validated by identification of several active inhibitors for Trypsin I Protease.
• Proof-of-concept applications: Demonstrates utility in de novo drug screening as a preliminary filtering step.

Scientific Applications:

• Early-stage virtual screening: Prioritizes compounds from large libraries for follow-up in drug discovery campaigns.
• Compound prioritization for docking/MD: Narrows large datasets to manageable subsets for docking and molecular dynamics simulations.
• De novo screening: Serves as an initial filter in de novo drug screening workflows.

Methodology:

Uses a deep learning-based Deep Feature Convolutional Neural Network for protein–ligand interaction prediction that produces prediction scores (e.g., with applied cutoffs such as 0.99).