Integrating Clinical, Pathologic, and Immune Features to Predict Breast Cancer Recurrence and Chemotherapy Benefit

摘要 乳腺癌是全球女性癌症死亡的主要原因，有超过 230 万病例 每年都会确诊。大多数病例为激素受体阳性，并可通过抗雌激素药物有效治疗 治疗，但有些患者患有侵袭性疾病，如果不进行治疗，就有复发和死亡的风险 化疗。基于基因表达的复发检测（例如 OncotypeDX）旨在预测 激素治疗复发并用于选择接受化疗的患者。然而，这些测定是 昂贵（每次测试 > 3,000 美元），需要相当长的时间才能进行，导致治疗延误，并且测试 在美国和全球资源匮乏的环境中，这些技术未得到充分利用或坦率地说不可用。相反，每个病人 患有乳腺癌的患者需要进行活检以确认诊断，病理学家会定期对活检进行分析以确定 乳腺癌的亚型和级别。深度学习是一种新兴的定量图像分析技术， 并且可以从病理学中识别非直观特征，包括基因表达模式。在前期工作中，我 已经证明，对病理样本的深度学习可以提供快速且经济有效的预测 OncotypeDX 评分使用现成的数据，可以识别激素复发风险较低的患者 治疗。 然而，OncotypeDX 仍然是化疗获益的不完美预测因子，因为它的开发目的是 预测激素治疗的复发。通过完善我的深度学习生物标志物，将临床和 乳腺癌的免疫特征，我可以提高化疗获益预测的准确性，从而提高 个性化治疗的能力。首先，我将利用国家最近扩大的临床数据 癌症数据库开发出更准确的预后和化疗益处的临床模型。接下来，我 将使用多重免疫荧光来更好地表征与相关的空间和细胞密度特征 化疗益处，并使用深度学习模型从标准苏木精和 伊红染色数字病理学。最后，我会将这些临床和免疫模型与我现有的深度结合起来 学习病理模型并在多机构队列中验证集成模型。这项工作的结果 将产生一种预后和预测性的深度学习生物标志物，可以轻松地做出准确的预测 可用的临床、病理和推断的免疫特征。这种方法有可能减少 由于周转时间过长而导致化疗延迟，通过提高可用性来消除医疗保健差异 测试，并通过定制用于预测化疗的生物标志物来提高治疗的个性化 好处。

Abstract Breast cancer is the leading cause of cancer death for women globally, with over 2.3 million cases diagnosed each year. Most cases are hormone receptor positive and effectively treated with anti-estrogen therapy, but some patients have aggressive disease and are at risk for recurrence and death without chemotherapy. Gene expression based recurrence assays, such as OncotypeDX, were designed to predict recurrence on hormonal therapy and are used to select patients for chemotherapy. However, these assays are expensive (> $3,000 per test), take considerable time to perform leading to treatment delays, and testing is underutilized or frankly unavailable in low resource settings in the US and globally. Conversely, every patient with breast cancer has a biopsy to confirm the diagnosis, which is routinely analyzed by pathologist to determine subtype of breast cancer and grade. Deep learning is an emerging technique for quantitative image analysis, and can identify non-intuitive features from pathology, including gene expression patterns. In preliminary work, I have demonstrated that deep learning on pathology samples can provide rapid and cost-effective prediction of OncotypeDX score using readily available data, and can identify patients at low risk of recurrence on hormonal therapy. However, OncotypeDX remains an imperfect predictor of chemotherapy benefit, as it was developed to predict recurrence on hormonal therapy. By refining my deep learning biomarker to incorporate clinical and immune features of breast cancer, I can improve accuracy in prediction of chemotherapy benefit and thus the ability to personalize treatment. First, I will capitalize on the recent expansion of clinical data in the National Cancer Data Base to develop a more accurate clinical models of prognosis and chemotherapy benefit. Next, I will use multiplex immunofluorescence to better characterize spatial and cell density features associated with chemotherapy benefit, and use deep learning models to infer these features from standard hematoxylin and eosin stained digital pathology. Finally, I will integrate these clinical and immune models with my existing deep learning pathologic model and validate the integrated model in a multi-institutional cohort. The result of this work will result in a prognostic and predictive deep learning biomarker that makes accurate predictions from readily available clinical, pathologic, and inferred immune features. This approach has the potential to reduce chemotherapy delays due to rapid turnaround time, combat healthcare disparities through improved availability of testing, and improve personalization of treatment by tailoring a biomarker for prediction of chemotherapy benefit.