利用深度神经网络分析ctDNA序列特征辅助肺癌的早期诊断

Lung cancer is one of the cancers with high morbidity and mortality in the world. Tissue biopsy is the standard diagnostic procedure for cancer. Biopsy may also provide material for genotyping, which can assist in the diagnosis and selection of targeted therapies but may fall short in cases of inadequate sampling, particularly from highly heterogeneous tumors. Traditional tissue biopsy suffers greater limitations in its prognostic capability over the course of disease, most obviously as an invasive procedure with potential complications, but also with respect to probable tumor clonal evolution and metastasis over time from initial biopsy evaluation. Recent work highlights circulating tumor DNA (ctDNA) present in the blood as a supplemental, or perhaps an alternative, source of DNA to identify the clinically relevant cancer mutational landscape. Indeed, this noninvasive approach may facilitate repeated monitoring of disease progression and treatment response, serving as a means to guide targeted therapies based on detected actionable mutations in patients with advanced or metastatic solid tumors. ctDNA was found to be positionally biased in tumor patients, suggesting that the ctDNA sequence pattern was different from the cfDNA sequence pattern released by normal cells. The research on the application of deep learning in medical diagnosis has become an international hot spot. Currently, there is no research on ctDNA sequence analysis based on deep learning. Based on the study of ctDNA mutation analysis in malignant tumors and auxiliary diagnosis of cancer using deep learning neural network, this project will carry out the study of cfDNA sequence characteristics in lung cancer. By constructing a cfDNA sequence database of lung cancer, the local features of cfDNA sequences were preliminarily identified and the composition differences between sequences were clarified. Using deep convolutional neural network and natural language processing model to learn ctDNA sequence characteristics, and to develop a set of deep learning system based on ctDNA to assist the early diagnosis of lung cancer and realize personalized precision medicine.

肺癌是全球发病率和死亡率较高的癌症之一，严重威胁人类生命和健康。组织活检作为诊断癌症的标准方法有其局限性。循环肿瘤DNA(ctDNA)能够捕捉肿瘤的全面信息，辅助癌症诊断和合理选择靶向治疗。研究发现ctDNA具有位置偏好性，表明ctDNA和其他循环游离DNA(cfDNA)的序列特征存在差异，为基于ctDNA序列特征进行肺癌早期诊断提供理论依据。深度学习在人工智能医疗诊断中的应用研究为当前国际热点，目前尚未发现基于深度学习分析ctDNA序列的研究。本项目将在前期开展的恶性肿瘤ctDNA突变分析和深度学习辅助诊断研究的基础上，开展肺癌患者ctDNA序列特征的研究。通过识别肺癌患者ctDNA序列的局部特征，明确序列间存在的组成差异；探索利用深度卷积神经网络和自然语言处理模型学习肺癌患者ctDNA序列特征，并以此为基础构建肺癌早期诊断模型，以期提高肺癌早期诊断率。

在所有癌症中，肺癌的发病率和死亡率均排在第一位，且容易发生转移，预后极差，早期很难被发现。因此，亟需一种有效的手段去识别肺癌风险高的患者，为早期诊断和提高患者生存率提供参考。现有研究表明循环肿瘤DNA(ctDNA)以非侵入性的方式能够在早期阶段展现恶性肿瘤的特征，实时监测肿瘤进展和治疗反应并指导有针对性的个体化临床治疗方案。其中，可将肿瘤相关偏好末端丰度作为ctDNA的一个可识别特征，区分开ctDNA与其他循环游离DNA(cfDNA)序列特征存在的差异，这为通过序列特征识别ctDNA早期肿瘤信号提供理论依据。本项目通过收集肺癌患者ctDNA和健康人群的cfDNA，识别肺癌患者ctDNA的序列特征，明确序列间存在的组成差异；利用基于Transformer的预训练自然语言模型BERT，和基于注意力机制的多实例学习算法，构建基于ctDNA特征的肺癌早期诊断模型；采用ROC曲线来评估模型区分肺癌患者和健康人群的能力，模型实现了AUC值0.9。表明模型学习到的特征可以区分癌症组和健康组。本项目使用的基于注意力机制的多实例学习算法不仅可以区分良恶性，更重要的是它可以提供每个序列对预测结果的贡献。建立基于ctDNA特征的肺癌早期诊断模型，为今后更广泛深入研究深度学习在肺癌中辅助早期诊断、提高检出率、追踪治疗效果提供研究基础。