Abstract Introduction: Cancer vaccine therapies rely on accurate personalized selection of immunizing peptides in order to potentiate tumor-specific immune responses against neoepitopes derived from somatic mutations. Given the unique accumulation of mutations in each tumor as well as the patient’s particular complement of HLA class I alleles, the ability to accurately predict which epitopes will be presented by tumor cells is a fundamental prerequisite for successful vaccine design. By utilizing a mono-allelic mass spectrometry (MS) strategy for profiling the endogenous HLA class I peptidome, we recently showed that prediction of endogenous presentation can be drastically improved when model training integrates peptide sequence along with intracellular signals such as likelihood of proteasomal processing and peptide abundance. Yet the limited set of mono-allelic data did not allow for deep comparative analysis across HLA- A, B, and C alleles, which can better inform pan-allele predictor design. Moreover, the significant variability in per-allele model performance remains unexplained. Methods: We recently developed a scalable mono-allelic MS technique to profile naturally presented peptides on HLA molecules, whereby the HLA class I deficient B721.221 cell line is transfected with HLA expression vectors coding for a single allele of interest and eluted HLA peptides are analyzed by LC-MS/MS. In addition, endogenously presented antigens on primary tumor-derived cell lines from 4 melanoma patients were also identified via MS. To extract knowledge from this unique dataset, we implemented computational tools to summarize, visualize, and compare the characteristics of HLA- A, B, C, and G alleles and developed a novel approach to define allele similarity that takes into account the collection of sub-motifs per allele. We trained neural network prediction models, validated their performance on internal and external datasets, and analyzed the variability in performance across alleles. Results: To date, we have generated binding data for 92 HLA- A, B, C and G alleles, identifying more than 190,000 peptides and covering the most frequent alleles in the population. Extensive mono-allelic profiling revealed that some alleles present non-9-mer peptides with high frequency. The availability of large number of non-9-mer peptides allowed us to build length-specific models that often performed better than the corresponding non-length-specific models currently used. We observe that HLA- A and B alleles present more peptides of length 10 and 11 than C alleles, while C alleles have a higher propensity for 8-mers. Correlation-based analysis of binding motifs revealed that HLA-A and B motifs are more specific whereas C motifs are less stringent and thus share more overlapping binders. Since binding data are available only for a fraction of all known alleles, pan-allele models implicitly embed allele similarity to predict for uncharacterized alleles based on the sequence of the binding pocket. By clustering allele-specific peptides into sub-motifs, we propose a novel explicit approach to delineate allele similarity at finer granularity that can improve pan-allele model design. We show that our allele-specific models are better at discriminating tumor-presented epitopes than state of the art predictors and investigate the relationship between false discovery rate and natural abundance of anchor residues to better understand differences in model accuracy amongst alleles. Finally, deconvolution of tumor-presented peptides demonstrated that ~10% of peptides are presented on HLA-C, which has been historically understudied. Conclusions: We have vastly expanded the collection of endogenous HLA-specific peptides deriving biologic insights into the principles of epitope presentations and valuable considerations for prediction model design and epitope selection for tumor vaccines. Citation Format: Siranush Sarkizova, Susan Klaeger, Derin B. Keskin, Karl Clauser, Hasmik Keshishian, Christina R. Hartigan, Nir Hacohen, Steven A. Carr, Catherine J. Wu. Broad analysis and more accurate predictions of HLA class I epitope binding in 92 common HLA alleles profiled by mono-allelic mass spectrometry [abstract]. In: Proceedings of the Fourth CRI-CIMT-EATI-AACR International Cancer Immunotherapy Conference: Translating Science into Survival; Sept 30-Oct 3, 2018; New York, NY. Philadelphia (PA): AACR; Cancer Immunol Res 2019;7(2 Suppl):Abstract nr B042.