TransformerCPI

TransformerCPI predicts compound–protein interactions (CPI) from protein sequence and compound information using a transformer neural network to enable CPI analysis for targets lacking resolved three-dimensional structural data.

Key Features:

• Sequence-Based Prediction: Uses protein sequences as the primary input, enabling CPI prediction for proteins without available structural data.
• Transformer Neural Network Architecture: Employs a transformer model with a self-attention mechanism to capture complex dependencies within sequence and compound representations.
• Label Reversal Experiments: Incorporates label reversal experiments to test that the model learns genuine interaction features rather than dataset artifacts or biases.
• Dataset Construction: Uses newly constructed datasets that address inappropriate datasets, hidden ligand bias, and improper dataset splitting to mitigate overestimation of model performance.
• Deconvolution Capability: Deconvolves model outputs to highlight critical interacting regions within protein sequences and compound atoms to inform ligand structural optimization.

Scientific Applications:

• Drug Discovery: Predicts CPIs to support identification and prioritization of potential therapeutic targets and compounds, including targets without resolved structures.
• Chemogenomics: Enables exploration of protein–ligand relationships across sequence space using sequence-only data.
• Ligand Optimization: Identifies important interacting protein regions and compound atoms to guide structural optimization of ligands.

Methodology:

Uses a transformer neural network with self-attention for sequence-based CPI prediction, performs label reversal experiments for validation, constructs datasets to avoid ligand bias and improper splits, and applies deconvolution of outputs to localize interacting residues and atoms.