Prediction of novel mouse TLR9 agonists using a random forest approach

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Toll-like receptor 9 (TLR9) is crucial for innate immunity against infections and cancer.
• TLR9 detects single-stranded DNA oligonucleotides (ODN) with CpG motifs.
• Traditional virtual screening of ODNs is challenging due to their structural complexity.

Purpose of the Study:

• Develop a machine learning (ML) method for predicting mouse TLR9 (mTLR9) agonists.
• Identify novel mTLR9 agonists using computational approaches.

Main Methods:

• Utilized an in-house dataset of 396 synthetic ODNs with experimental TLR9 activity data.
• Employed an ensemble learning approach with repeated random down-sampling to address data imbalance.
• Compared five ML algorithms, focusing on Random Forest for its superior performance.

Main Results:

• The Random Forest ensemble classifier achieved an average balanced accuracy of 80% and MCC of 0.61.
• Maximum balanced accuracy reached 87% and MCC 0.75 in test samples.
• Experimental validation of top predictions showed 91% of 100 synthesized ODNs were active mTLR9 agonists.

Conclusions:

• Ensemble Random Forest with down-sampling effectively predicts mTLR9 agonists, overcoming class imbalance.
• Random Forest outperformed SVM, SDA, GBM, and Neural Networks for this prediction task.
• This ML approach offers a robust and simple method for identifying potential mTLR9 ODN agonists.