Abstract 3210: Strategies to improve the sensitivity and ranking ability of neoantigen prediction methods: Report on the results of the Tumor nEoantigen SeLection Alliance (TESLA)

Neoantigen-based therapies hold the promise of being safe, personalized anti-cancer therapies for a broad range of cancers. A crucial step for the development of effective neoantigen therapies is the identification of putative tumor specific neoantigens typically achieved through in silico analyses. In silico-based approaches have the benefit of being fast, modular and reproducible. A wide range of tools are available for variant calling, determining peptide-HLA binding affinity, peptide foreignness, agritopicity and more, and researchers are continuously adding to the lot. However, there does not currently exist a common reference data set with which these different approaches can be compared, and the key parameters for effective neoantigen identification remain elusive. Here we introduce a consortium-based initiative, the Tumor nEoantigen SeLection Alliance (TESLA), to address these needs and describe straight-forward strategies through which the sensitivity and ranking ability of neoantigen prediction methods can be improved. TESLA participants were provided with whole exome and RNA sequencing from six tumors (3 melanoma and 3 NSCLC) and in turn submitted predictions for immunogenic peptides. Twenty-five groups from academia, pharma and biotech participated to this challenge. Peptides identification and ranking across the teams were assessed and showed little overlap. The limited overlap could not be explained simply by differences in variant calling. From these predictions, 608 peptides regrouping the teams' highly ranked epitopes plus epitopes frequently called and ranked across the teams, were assessed for T-cell recognition in patient-matched samples as well as for in-vitro MHC binding affinity. Analyses showed that peptides recognized by T-cells had significantly stronger MHC binding affinity and were derived from genes with significantly higher gene expression. Concomitantly, neoantigen pipelines which prioritized epitopes with strong binding affinity and/or which filtered out those originating from genes with low tumor variant allele fraction or low gene expression were associated with improved ability to identify and rank neoantigens. Direct interventions on participant predictions using these identified traits demonstrated substantial improvement in the performance of the neoantigen predictions. Pipeline analysis indicated that there are a range of approaches to neoantigen prediction and improving upon an existing neoantigen pipeline requires assessing that pipeline for a range of characteristics, including variant detection ability, peptide filtering ability, and peptide ranking ability, and improving those areas where the pipeline performance is sub-optimal. TESLA data will be continually available and serve as a living benchmark to improve neoantigen prediction. Citation Format: Daniel K. Wells, Kristen Dang, Vanessa M. Hubbard-Lucey, Kathleen C. Sheehan, Andrew Lamb, Jeffrey P. Ward, John Sidney, Ana B. Blazquez, Andrew J. Rech, Jesse Zaretsky, Begonya Comin-Anduix, Alphonsus H. Ng, William Chour, Thomas V. Yu, Hira Rizvi, Jia Chen, Patrice Manning, Taha Merghoub, Justin Guinney, Adam Kolom, Cheryl Selinsky, Antoni Ribas, Matthew D. Hellmann, Ton N. Schumacher, Nir Hacohen, Pia Kvistborg, Alessandro Sette, James R. Heath, Nina Bhardwaj, Fred Ramsdell, Robert D. Schreiber, Nadine A. Defranoux, TESLA Consortium. Strategies to improve the sensitivity and ranking ability of neoantigen prediction methods: Report on the results of the Tumor nEoantigen SeLection Alliance (TESLA) [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 3210.