Improved cytokine-receptor interaction prediction by exploiting the negative sample space.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Cytokine-receptor interactions (CRIs) are crucial for understanding human diseases like autoimmune and inflammatory conditions.
• Accurate prediction of CRIs is vital for identifying therapeutic targets.
• Existing machine learning models for CRIs suffer from biases in negative datasets, impacting training and evaluation.

Purpose of the Study:

• To address biases in negative sample selection for cytokine-receptor interaction prediction.
• To propose and evaluate a clustering-based approach for representative negative sample selection.
• To improve the accuracy and reliability of machine learning models for CRIs.

Main Methods:

• Utilized deep autoencoders to analyze the impact of different negative sampling strategies.
• Implemented K-means based sampling to create representative negative datasets.
• Employed random forest (RF) classifiers with a combined feature set including atomic composition, physicochemical-2grams, and evolutionary information.

Main Results:

• K-means based sampling significantly outperformed random sampling in mitigating negative dataset biases.
• Random forest models trained on K-means sampled data showed superior performance: +5.1% accuracy, +13% specificity, +0.1 MCC, and +5.1% g-means.
• Deep autoencoders revealed how different negative sample categories affect learning algorithm training.

Conclusions:

• K-means sampling provides a more representative negative dataset, leading to improved supervised learning model performance in bioinformatics.
• Random forest models demonstrated the most significant benefits from K-means sampled datasets for CRI prediction.
• The proposed sampling methodology is highly relevant for various bioinformatics applications relying on supervised learning.