Abstract
Neoepitope-based cancer immunotherapy depends on accurate prediction of patient-specific neoepitopes. Many candidate neoepitopes can be identified but their prioritization is challenging, resulting in poor effectiveness of existing methods. NeoGuider, our neoepitope prediction pipeline, detects neoepitope candidates from sequencing data and utilizes machine learning to prioritize and probabilistically classify the candidates to address the challenges, we developed a novel feature transformation in NeoGuider which uses adaptive kernel density estimation and centered isotonic regression to transform feature values into log odds. We studied the performance of NeoGuider on six cohorts, encompassing 43 patients with 168 immunogenic candidates. Experiments showed that it outperformed existing methods. NeoGuider is open-sourced at https://github.com/XuegongLab/neoguider.