Background: Immune checkpoint inhibitors (ICIs) have demonstrated significantly improved clinical efficacy in a minority of patients with advanced melanoma, whereas non-responders potentially suffer from severe side effects and delays in other treatment options. Predicting the response to anti-PD1 treatment in melanoma remains a challenge because the current FDA-approved gold standard, the nonsynonymous tumor mutation burden (nsTMB), offers limited accuracy. Methods: In this study, we developed a multi-omics-based machine learning model that integrates genomic and transcriptomic biomarkers to predict the response to anti-PD1 treatment in patients with advanced melanoma. We employed least absolute shrinkage and selection operator (LASSO) regression with 49 biomarkers extracted from tumor–normal whole-exome and RNA sequencing as input features. The performance of the multi-omics AI model was thoroughly compared to that of nsTMB alone and to models that use only genomic or transcriptomic biomarkers. Results: We used publicly available DNA and RNA-seq datasets of melanoma patients for the training and validation of our model, forming a meta-cohort of 449 patients for which the outcome was recorded as a RECIST score. The model substantially improved the prediction of anti-PD1 outcomes compared to nsTMB alone, with an ROC AUC of 0.7 in the training set and an ROC AUC of 0.64 in the test set. Using SHAP values, we demonstrated the explainability of the model’s predictions on a per-sample basis. Conclusions: We demonstrated that models using only RNA-seq or multi-omics biomarkers outperformed nsTMB in predicting the response of melanoma patients to ICI. Furthermore, our machine learning approach improves clinical usability by providing explanations of its predictions on a per-patient basis. Our findings underscore the utility of multi-omics data for selecting patients for treatment with anti-PD1 drugs. However, to train clinical-grade AI models for routine applications, prospective studies collecting larger melanoma cohorts with consistent application of exome and RNA sequencing are required.
背景：免疫检查点抑制剂（ICIs）在少数晚期黑色素瘤患者中显示出显著改善的临床疗效，而无应答者则可能遭受严重副作用并延误其他治疗选择。预测黑色素瘤对抗PD1治疗的反应仍具挑战性，因为目前FDA批准的金标准——非同义肿瘤突变负荷（nsTMB）的准确性有限。方法：本研究开发了一种基于多组学的机器学习模型，整合基因组和转录组生物标志物，以预测晚期黑色素瘤患者对抗PD1治疗的反应。我们采用最小绝对收缩和选择算子（LASSO）回归，以从肿瘤-正常全外显子组和RNA测序中提取的49个生物标志物作为输入特征。该多组学AI模型的性能与单独使用nsTMB以及仅使用基因组或转录组生物标志物的模型进行了全面比较。结果：我们利用公开可用的黑色素瘤患者DNA和RNA-seq数据集进行模型的训练和验证，形成了一个包含449名患者的元队列，其治疗结果以RECIST评分记录。与单独使用nsTMB相比，该模型显著改善了对抗PD1治疗结果的预测，在训练集中的ROC AUC为0.7，在测试集中的ROC AUC为0.64。通过使用SHAP值，我们展示了模型在单样本基础上预测的可解释性。结论：我们证明，仅使用RNA-seq或多组学生物标志物的模型在预测黑色素瘤患者对ICI的反应方面优于nsTMB。此外，我们的机器学习方法通过提供针对每位患者的预测解释，提高了临床可用性。我们的研究结果强调了多组学数据在筛选适合接受抗PD1药物治疗患者方面的实用性。然而，为训练适用于常规临床应用的AI模型，仍需开展前瞻性研究，收集更大规模的黑色素瘤队列，并一致应用外显子组和RNA测序技术。
AI Model for Predicting Anti-PD1 Response in Melanoma Using Multi-Omics Biomarkers
肿瘤(癌症)患者之家