Abstract B056: Deep Antigen Sets, a new deep-learning framework, enables in vivo modeling of patient neoantigens for effective targeted immunotherapy design

Abstract Background: Personalized cancer vaccines hold immense promise, yet their efficacy hinges on accurately predicting immunogenic neoantigens—tumor-specific peptides that elicit robust immune responses. Current machine learning algorithms, trained using in vitro epitope-level immunogenicity data, are limited by the scarcity of experimentally validated immunogenic epitopes and often fail to generalize to immune responses observed in vivo. Methods: We introduce Deep Antigen Sets (DAS), a novel deep-learning framework that enhances neoantigen prediction by directly modeling patient-level immune responses while making precise predictions at the individual neoantigen level. Using DeepSets, a permutation-invariant architecture, DAS models the set of potential neoantigens, which will vary in size, as a holistic input. This allows us to uniquely leverage patient-level clinical outcomes, bridging the gap between in vitro and in vivo predictions. Furthermore, we implemented a targeted sampling strategy to filter out likely non-immunogenic epitopes, refining the training set with the most informative neoantigen candidates to enhance robustness and generalization. DAS is evaluated on two fronts: first, by assessing its ability to prioritize experimentally validated immunogenic neoantigens among the top-ranked predictions for each patient, and second, by predicting whether a patient is likely to respond to immunotherapy based solely on their neoantigen profile. Results: The performance of DAS was benchmarked on three external clinical datasets—NCI, TESLA, and HiTIDE—each containing experimentally validated immunogenic neoantigens, as well as on an internal test cohort annotated with patient responses to immunotherapy. After training on neoantigen sets from 543 in-house patients, DAS achieved 69% AUC in predicting clinical response on the internal hold-out cohort. Furthermore, DAS consistently outperformed existing neoantigen prediction models across all three external clinical datasets. Specifically, it recovered over 60%, 80%, and 90% of validated immunogenic neoantigens within the top 20, 50, and 100 ranked predictions per patient, respectively. Conclusions: DAS constitutes a paradigm shift by directly modeling in vivo neoantigen sets and patient responses to enhance the potential of personalized cancer vaccines. Its computational efficiency facilitates the rapid analysis of large patient cohorts, seamlessly integrating into clinical workflows. This powerful tool empowers oncologists and immunologists to design more effective targeted immunotherapies, ultimately improving patient outcomes. Citation Format: Michail Chatzianastasis, Robert Burns, Helio Costa, Matthew Rabinowitz, Frank Zhang. Deep Antigen Sets, a new deep-learning framework, enables in vivo modeling of patient neoantigens for effective targeted immunotherapy design [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B056.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.