DeepPROTACs

DeepPROTACs predicts PROTAC degradation efficacy using deep learning models to support rational design of targeted protein degraders by integrating structural data of target proteins and E3 ligases.

Key Features:

• Deep Neural Network Model: The core architecture leverages both graph-based and sequential data representations to predict PROTAC degradation potential.
• Graph Convolutional Networks (GCNs): Ligands and their binding pockets are modeled as graphs and processed with GCNs for molecular feature extraction.
• Bidirectional Long Short-Term Memory (BiLSTM): SMILES representations of linkers are fed into a BiLSTM layer to capture sequential dependencies in linker sequences.
• Experimental Dataset from PROTAC-DB: Training and validation use experimental entries sourced from PROTAC-DB labeled by DC50 and Dmax values.
• Performance Metrics: Reported predictive performance includes an average accuracy of 77.95% and an AUC-ROC of 0.8470.

Scientific Applications:

• Rational PROTAC design: Predicts degradation capacity to assist selection and optimization of PROTAC molecules.
• Targeted protein degradation and drug discovery: Prioritizes candidate PROTACs for further experimental evaluation in therapeutic development.

Methodology:

Integrates structural data of target proteins and E3 ligases with molecular features from ligands and linkers; uses GCNs for graph-based feature extraction and BiLSTM layers for sequential SMILES processing; training and validation are performed on PROTAC-DB data labeled by DC50 and Dmax.