Abstract Background: Personalized neoantigen-based therapies are being actively developed with several clinical trials currently ongoing. The standard methodology to identify patient specific neoantigens relies on computational algorithms for analyzing mutation, expression, and HLA profiles from patients’ NGS data. However, clinical reports have thus far showed that the number of actionable neoantigens was not as high as originally expected. One possible reason is due to a lack of appropriate prediction models, especially for immunogenicity, which is largely hindered by a scarcity of training data based on in vivo vaccination models. In addition, HLA allele status and/or impurity and heterogeneity in tumor samples are other factors affecting the selection efficiency. Methods: We have developed a computational pipeline to identify neoantigens from patients’ WES and RNA-Seq data. Our HLA assessment not only determines HLA types but also detects HLA LOH events from WES and down-regulation of HLA gene expression from RNA-Seq taking into account tumor purity. The neoantigen candidates were ranked based on a presentation score calculated by a regression model optimized with publicly available immunopeptidome data. We tested our top-ranked neoantigen peptides by immunizing them into HLA-A02:01, A24:02, B07:02, and B35:01-transgenic mice and performed ELISPOT to determine positive or negative in vivo immune responses. Furthermore, we trained a neural network to define an immunogenicity score based on the results from the transgenic mouse experiments. Results: We generated NGS data from over 100 clinical samples and analyzed the data using our computational pipeline. We observed a wide range in the tumor mutational burden and tumor purity of our samples, and HLA LOH was frequently observed. From 27 patient cases, 275 of HLA-A02:01-, A24:02-, B07:02-, or B35:01-predicted neoantigen epitopes were tested in transgenic mice and confirmed that the rate of positively reacted epitopes increased in a correlated manner to the presentation score, achieving 82% positive rate in top scored 35 epitopes and 47% in overall 275 epitopes. We further developed an immunogenicity prediction model and tested its performance in additional clinical samples. By integrating it into the selection criteria, we achieved higher positive rate than using the presentation score solely. Conclusions: We have developed a robust computational pipeline to identify putatively actionable neoantigens from patients’ WES and RNA-Seq data. We detected frequent HLA LOH events indicating that this is a common mechanism of immune evasion. We also showed the feasibility of improving prediction algorithms using HLA-transgenic mice. Accurate prediction performance enables us to decrease the number of vaccine antigens to be immunized per patient, leading to a decrease in the manufacturing cost and complexity in quality control. Citation Format: Kazushi Hiranuka, Keigo Takahashi, Dave Tang, Noriko Watanabe, Takashi Yamada, Yuji Mishima, Norihiro Fujinami, Manami Shimomura, Toshihiro Suzuki, Tetsuya Nakatsura, Norihiro Nakamura. A novel computational pipeline supported with in vivo vaccination models for identification and validation of neoantigens [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 6389.