iTTCA-RF: a random forest predictor for tumor T cell antigens.

Area of Science:

• Oncology
• Immunology
• Bioinformatics

Background:

• Cancer immunotherapy offers a promising treatment strategy with high efficacy and fewer side effects.
• Accurate identification of tumor T cell antigens is vital for developing antitumor vaccines and investigating molecular functions.
• Existing machine learning methods for tumor T cell antigen identification face challenges in accuracy.

Purpose of the Study:

• To develop a more accurate computational tool for identifying tumor T cell antigens.
• To improve the prediction of tumor T cell antigens using advanced feature engineering and machine learning algorithms.

Main Methods:

• Utilized a non-redundant dataset comprising 592 positive and 393 negative tumor T cell antigen samples.
• Employed four feature encoding methods: amino acid composition, global protein sequence descriptors, grouped amino acid, and peptide composition.
• Implemented a two-step feature selection technique and a random forest algorithm to build the prediction model.

Main Results:

• Selected the top 263 informative features to train the random forest classifier.
• Achieved balanced accuracy of 83.71% (10-fold CV) and 73.14% (independent tests).
• Demonstrated high sensitivity (88.69% 10-fold CV, 83.61% independent tests) and specificity (78.73% 10-fold CV, 62.67% independent tests).

Conclusions:

• The developed predictor, iTTCA-RF, shows superior performance compared to existing models.
• iTTCA-RF is a valuable tool for identifying tumor T cell antigens, particularly those presented in the context of major histocompatibility complex class I.
• An online prediction server is available for public use.