通过进化选择压力预测新抗原免疫原性

Introduction: Accurate prediction of neoantigen immunogenicity remains a critical challenge in cancer immunotherapy and personalized vaccine design. Current computational approaches show limited predictive power, with existing methods achieving modest performance in identifying truly immunogenic neoantigens. We developed NEMo (Neoantigen Elimination Model), an evolutionary machine learning model that predicts neoantigen immunogenicity by learning mutation characteristics that render neoantigens particularly visible to the immune system and are subsequently eliminated from tumors under immune selective pressure following immune checkpoint blockade (ICB) treatment. Experimental Procedures: NEMo was trained on over 25,000 mutations with longitudinal time-point data from patients treated with immune checkpoint blockade. The model leverages the selective pressure exerted by the immune system-whereby immunogenic neoantigens are preferentially shed from tumors after treatment-as a potent proxy for immunogenicity. We validated NEMo's performance across four independent datasets: two assessing general neoantigen immunogenicity and two from cancer vaccine clinical trials. Results: In a cohort of 74 gastric cancer patients, NEMo distinguished immunogenic from non-immunogenic neoantigens with area under the curve (AUC) values of 0.855 for CD8+ T cell reactivity and 0.73 for CD4+ T cell reactivity (P=2e-17 and P=4e-10, respectively). In the TESLA consortium global competition involving 25 teams, NEMo substantially outperformed competing methods with 74% accuracy compared to 42% for the next best competitor (average 15%). In the Ott et al. melanoma vaccine trial (n=8 patients), NEMo correctly identified immunogenic neoantigens with AUCs of 0.87 for CD8+ and 0.7 for CD4+ responses. Similar performance was observed in the Cafri et al. vaccine trial of colorectal and gastric cancer patients (n=4), with AUCs of 0.7 for both CD8+ and CD4+ responses. Notably, NEMo's superior performance can be attributed to enhanced modeling of HLA presentation dynamics. In the Ott dataset, seven neoantigens originally selected as highly immunogenic candidates by trial designers were well-presented by HLA in silico but showed reduced immunogenic potential due to tumor evolution-mediated deletion of specific HLA alleles (P<0.001). Neoantigens unaffected by these deletions demonstrated significantly higher immunogenicity (P<0.001). Conclusions: NEMo represents a recent and groundbreaking advancement in computational neoantigen immunogenicity prediction, leveraging evolutionary selection pressure as a biologically grounded training signal, free from in-vivo distortions to the immune system. By accurately modeling HLA presentation and incorporating tumor evolution dynamics, NEMo provides a powerful tool for rational design of cancer vaccines and patient stratification for immunotherapy.