A New Estimation Method for the Biological Interaction Predicting Problems.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Computational methods are crucial for predicting biological interactions.
• Existing methods often treat prediction as semi-supervised or positive-unlabeled (PU) learning.
• A key limitation is the inability to estimate the total number of undiscovered interactions within datasets.

Purpose of the Study:

• To develop a deep learning-based estimation method for quantifying undiscovered biological interactions in unlabeled samples.
• To provide asymptotic interval estimation for the calculated number of undiscovered interactions.
• To offer guidance for experimental validation by identifying datasets with higher potential for novel interactions.

Main Methods:

• Developed a novel deep learning model for interaction prediction and estimation.
• Implemented asymptotic interval estimation for the model's predictions.
• Applied the method to compound synergism, drug-target interaction (DTI), and microRNA-disease interaction datasets.
• Compared the proposed method with existing mixture proportion estimators.

Main Results:

• Successfully estimated the number of undiscovered interactions across multiple biological datasets.
• Demonstrated the method's ability to identify datasets with a higher prevalence of novel interactions.
• Showcased the efficacy of the deep learning approach compared to traditional estimators.
• Established a relationship between AUC/AUPR and the number of undiscovered interactions, proposing them as new evaluation metrics.

Conclusions:

• The developed deep learning method accurately estimates undiscovered biological interactions.
• This estimation provides valuable guidance for prioritizing experimental validation efforts.
• The findings offer a new perspective on evaluating computational interaction prediction methods using AUC and AUPR.