DeepZF

DeepZF predicts which zinc-finger (ZF) domains within Cys2His2 (C2H2) zinc-finger proteins bind specific DNA triplets and infers their DNA-binding preferences from ZF amino-acid sequences.

Key Features:

• Binding prediction: Predicts binding ZFs and DNA-binding preferences from amino-acid sequences of ZFs.
• Model architecture: Employs a novel protein transformer deep-learning model.
• Training data: Trained on an in vivo dataset comprising both binding and non-binding ZFs.
• Classifier component: Implements a ZF-binding classifier that distinguishes binding versus non-binding ZFs with an average AUROC of 0.71.
• Preference predictor: Implements a DNA-binding preference predictor that uses transfer learning integrating in vivo and in vitro datasets.
• Prediction performance: Achieves average Pearson correlations >0.94 across each of the three DNA binding positions for DNA-binding predictions.
• Comparative performance: Outperforms existing methods with reported motif similarity correlation of 0.42 versus correlations typically below 0.1 for prior methods.
• Interpretability: Applies established interpretability techniques to relate specific amino-acid residues to ZF DNA-binding potential.

Scientific Applications:

• Gene regulation studies: Facilitates analysis of C2H2-ZF transcription factor DNA-recognition and binding specificity.
• Disease and cellular function research: Supports investigation of cellular functions and disease mechanisms involving C2H2-ZF proteins.
• Synthetic biology: Aids design and engineering of ZF-based DNA-binding modules through precise DNA-binding predictions.

Methodology:

Uses a protein transformer deep-learning model trained on an in vivo dataset of binding and non-binding ZFs; implements a ZF-binding classifier (AUROC 0.71) and a DNA-binding preference predictor trained via transfer learning on combined in vivo and in vitro datasets (Pearson correlations >0.94 across three DNA binding positions); applies interpretability techniques to connect amino-acid residues to binding preferences.