Yan Kai, Zhou Zhiheng, Liu Sihao, Wang Guanghui, Yan Guiying, Wang Edwin
IEEE J Biomed Health Inform. 2025 Mar 28;PP. doi: 10.1109/JBHI.2025.3555596.
The emergence of immune checkpoint inhibitors (ICIs) has significantly advanced cancer treatment. However, only 15-30% of the cancer patients respond to ICI treatment, which stimulates and enhances host immunity to eliminate tumor cells. ICI treatment is very expensive and has potential adverse reactions; therefore, it is crucial to develop a method which enables to accurately and rapidly assess a patient's suitability before ICI treatment. We complied germline whole-genome sequencing (WES) data of 37 melanoma patients who have been treated with ICIs and sequenced in our lab previously, and the WES data of other 700 ICI-treated cancer patients in public domain. Using these data, we proposed a novel double-channel attention neural network (DANN) model to predict cancer ICI-response and validate the predictions. DANN achieved a mean accuracy and AUC of 0.95 and 0.98, respectively, which outperformed traditional machine learning methods. Enrichment analysis of the DANN-identified genes indicated that cancer patients whose in-born genomic variants might mainly affect host immune system in a wide-ranging manner, and then affect ICI response. Finally, we found a set of 12 genes bearing genomic variants were significantly associated with cancer patient survivals after ICI treatment.
免疫检查点抑制剂（ICI）的出现显著推动了癌症治疗的发展。然而，只有15%至30%的癌症患者对ICI治疗有反应，ICI治疗通过刺激和增强宿主免疫力来消除肿瘤细胞。ICI治疗费用高昂且存在潜在不良反应；因此，开发一种能够在ICI治疗前准确、快速评估患者适用性的方法至关重要。我们整理了37例曾接受ICI治疗且此前在我们实验室进行过测序的黑色素瘤患者的种系全基因组测序（WES）数据，以及公开领域中其他700例接受ICI治疗的癌症患者的WES数据。利用这些数据，我们提出了一种新型双通道注意力神经网络（DANN）模型来预测癌症对ICI的反应并验证预测结果。DANN的平均准确率和AUC分别达到了0.95和0.98，优于传统机器学习方法。对DANN识别出的基因进行的富集分析表明，先天基因组变异可能主要以广泛的方式影响宿主免疫系统，进而影响ICI反应。最后，我们发现一组携带基因组变异的12个基因与ICI治疗后癌症患者的生存率显著相关。
