Prediction of bacterial protein-compound interactions with only positive samples

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology
• Drug Discovery

Background:

• Accurate prediction of Compound-Protein Interactions (CPI) is vital for pharmaceutical and chemical engineering applications.
• Existing CPI models are inadequate for bacterial systems due to a lack of negative interaction data.
• Bacterial cytochrome P450 (CYP) interactions are of particular interest.

Purpose of the Study:

• To develop a novel framework for predicting bacterial Compound-Protein Interactions (CPI).
• To address the challenge of limited negative samples in bacterial CPI prediction.
• To enhance the accuracy and applicability of deep learning models for CPI prediction in bacteria.

Main Methods:

• Proposed a Positive-Unlabeled (PU) learning framework named BIN-PU.
• Implemented a strategy to generate pseudo positive and negative labels from known positive interactions.
• Developed a weighted positive loss function to prioritize truly positive samples.
• Validated BIN-PU with various CPI backbone models on bacterial CYP data.

Main Results:

• BIN-PU demonstrated superior performance in predicting CPIs using only positive samples compared to existing PU models.
• The framework showed reproducibility across diverse bacterial protein datasets, including human CYP and uncurated data.
• Experimental validation confirmed the accuracy of BIN-PU's CPI predictions for uncurated CYP data.

Conclusions:

• BIN-PU offers a significant advancement for predicting bacterial Compound-Protein Interactions (CPI).
• The framework enhances the predictive power of deep learning models in biological interaction tasks.
• BIN-PU opens new avenues for research in drug discovery and biocatalysis involving bacterial proteins.