Title: Bioinformatics and AI methods for neoepitope prediction in personalized cancer immunotherapy
PhD Candidate: Elham Sherafat
Major Advisor: Dr. Ion Mandoiu
Associate Advisors: Dr. Mukul Bansal, Dr. Sheida Nabavi
Day/Time: Tuesday, November 26, 2019 3:00-4:00 PM
Location: ITE 336
Personalized cancer vaccines are emerging as one of the most promising approaches for immunotherapy of advanced cancers. This approach harnesses the power of the patient’s own immune system to attack tumor cells that express specific neoepitopes generated by somatic mutations. Calling somatic variants from matched tumor-normal next-generation sequencing (NGS) data is a crucial step in the identification of neoepitopes that can be included in a cancer vaccine. Although many somatic variant callers exist based on a variety of statistical models, the agreement between different callers is low, and accurate somatic variant calling remains challenging. Key impediments to achieving consistently high accuracy with model-based methods include the large patient-to-patient variation in sample attributes such as purity, tumor heterogeneity, sequencing library preparation artifacts, sequencing errors, and errors in NGS data processing such as incorrect read alignment.
In this proposal, we explain a novel machine learning method to increase the performance of any existing somatic variant calling pipeline while maintaining high positive predictive value. To reliably handle patient-to-patient variation in sample attributes, we take a semi-supervised approach that learns these properties from the data itself, without a need for prior training data. Experimental results on cancer sequencing data from ovarian cancer patients enrolled in an ongoing Phase I clinical trial at UConn Health show the effectiveness of our approach. We also show that our method improves the rate of confident peptide identification compared to existing methods for analyzing mass spectra generated from MHC-I eluted peptides.
In ongoing work, we are exploring supervised and semi-supervised methods for predicting neoepitope immunogenicity. This is a crucial task since it is impractical to experimentally assess all candidate neoepitopes prior to vaccination.