imPlatelet classifier: Image-converted RNA biomarker profiles enable blood-based cancer diagnostics

Abstract: Sequencing technologies have enabled in-depth analysis of liquid biopsies in cancer, offering a minimally invasive sample collection. The most widely used material is blood which, next to circulating tumor cells and circulating tumor DNA, is the source of tumor-educated platelets (TEPs). Methods: We developed imPlatelet method which converts RNA-sequenced platelet data to images, additionally implementing biological knowledge from the Kyoto Encyclopedia of Genes and Genomes Pathway. First, we tested imPlatelet method on a cohort of 401 non-small cell lung cancer patients and 62 sarcoma patients. Next, we applied the developed tool to platelets collected from a new, independent cohort of 28 ovarian cancer patients and 30 non-cancer benign gynaecological conditions. Results: imPlatelet provided excellent discrimination between cancer cases and healthy controls, with accuracy equal to 1 in training, validation and independent datasets. When discriminating between ovarian cancer cases and benign conditions, imPlatelet reached 0.91 balanced accuracy, with sensitivity and specificity equal to 0.95 and 0.88, respectively, in an independent test set. ImPlatelet outperformed current state-of-the-art method thromboSeq in the aspects of balanced classification accuracy, the computational power needed, user experience, and execution time. Conclusions: According to our knowledge, this is the first study implementing an image-based deep learning approach combined with biological knowledge to classify human samples. Our results on classification of ovarian cancer considerably outperform previously published methods and our own alternative attempts of discrimination. We show that a deep learning image-based classifier accurately identifies cancer, despite the limited number of samples and even among non-cancer conditions which affect platelet transcriptome making the diagnosis difficult. RNA-seq of Tumor-educated platelets in ovarian carcinoma

摘要：测序技术推动了癌症液体活检的深度分析，为样本采集提供了微创途径。目前应用最广泛的样本材料为血液，除循环肿瘤细胞（circulating tumor cells）和循环肿瘤DNA（circulating tumor DNA）外，血液也是肿瘤教育血小板（tumor-educated platelets, TEPs）的来源。

方法：我们开发了imPlatelet方法，该方法可将血小板RNA测序数据转化为图像，并整合了来自京都基因与基因组百科全书通路（Kyoto Encyclopedia of Genes and Genomes Pathway）的生物学知识。首先，我们在包含401例非小细胞肺癌患者与62例肉瘤患者的队列中对imPlatelet方法进行了测试。随后，我们将该开发工具应用于新的独立队列采集的血小板样本，该队列包含28例卵巢癌患者与30例非癌性良性妇科疾病患者。

结果：imPlatelet可精准区分癌症病例与健康对照，在训练集、验证集与独立数据集上的准确率均达到1。在区分卵巢癌病例与良性妇科疾病的任务中，imPlatelet在独立测试集上的平衡准确率达到0.91，敏感性与特异性分别为0.95与0.88。imPlatelet在平衡分类准确率、所需计算资源、用户体验与运行时长方面均优于当前最优方法thromboSeq。

结论：据我们所知，本研究是首个将基于图像的深度学习方法与生物学知识相结合以对人类样本进行分类的研究。我们在卵巢癌分类任务中得到的结果显著优于已发表的相关方法以及我们此前进行的其他分类尝试。研究表明，即便样本量有限，且存在影响血小板转录组从而增加诊断难度的非癌性疾病干扰，基于图像的深度学习分类器仍可精准识别癌症。

卵巢癌肿瘤教育血小板的RNA测序