常壓游離質譜法結合機器學習於乳癌診斷之應用

The Application of Ambient Ionization Mass Spectrometry and Machine Learning in Breast Cancer Diagnosis

黃映晨(Ying-Chen Huang)

指導教授 ： 徐丞志

國立臺灣大學/理學院/化學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201900343

摘要

近十年來，一系列常壓游離質譜法 (ambient ionization mass spectrometry) 的快速發展，已經被廣泛引入生物醫學領域之應用，並給癌症檢測提供了一個新的有效分析工具。樣品無須經過繁雜的前處理以及層析管柱分離為常壓游離質譜法的特點，生物樣品在常溫常壓下直接將表面的分析物脫附且同時游離後，進入到質譜儀中被偵測，大幅縮短分析所需時間，其中電噴灑離子化 (Electrospray ionization, ESI) 技術開發出的紙噴灑游離法 (paper spray ionization, PSI)以及脫附電灑游離法 (Desorption Electrospray Ionization, DESI)，目前已被大量應用於醫學檢驗以及癌症之快速診斷中。
惡性腫瘤，亦即癌症，為世界大部份先進國家的頭號殺手，其中女性罹癌的人口中，以乳癌的比例最高，其發生率在台灣已達到每十萬名女性有70人罹患乳癌，可說是威脅女性生命最高的疾病之一。本論文中，採用紙噴灑游離法 (PSI) 搭配軌道離子阱質譜儀（Orbitrap）應用於乳癌診斷，在醫院的常規檢查中，粗針穿刺切片為醫院常用之取樣方式，醫師通常以皮下針取患部組織，並將組織沾黏於濾紙上，作為後續病理學檢驗之樣本。以該含有組織的紙片樣品，於紙片電噴灑游離法下進行代謝物質分布的快速檢測，判斷該組織為癌症/非癌症。為了降低來自複雜基質的化學干擾，於紙片電噴灑游離源後方裝設場不對稱離子遷移譜（FAIMS-PSI-MS）。因此，可以有效地獲得粗針穿刺切片組織中預測性脂質圖譜， 透過機器學習演算法搭配病理學檢查，開發一個以小分子代謝物和脂質圖譜作為分類依據的模型，其可快速區分癌/非癌症乳腺組織。
此外，根據乳癌細胞上的賀爾蒙受體的陰陽性，如動情激素接受體 (ER)、黃體激素接受體 (PR)、以及第二型人類上皮細胞生長素接受體 (Her-2)，乳癌又可分為四種亞型，包含管腔細胞A型、管腔細胞B型、HER2過度表達型以及三陰性乳癌，乳癌細胞含有越多的荷爾蒙受體，它的治療效果越好，預後較佳。因此，乳癌一定要檢測荷爾蒙受體的含量，才能決定後續的治療方法。本論文中以DESI質譜影像技術，直接取得乳癌手術冷凍切片的脂質組成、以及小分子代謝物的化學資訊，每個DESI質譜影像的pixel會被視為一個資料點，並且以病理師對H&E染色的結果判讀為這些資料點分別標上「癌症」或是「正常」標籤，來自不同亞型的乳癌組織中，癌化與正常組織的質譜訊號會作為訓練集，並用不同機器學習的演算法如貝氏分類器、支持向量基、套索迴歸分析、決策樹、隨機森林或是深度學習法等，來訓練監督式學習的機器學習模型。藉由DESI質譜影像法所得數據結合機器學習演算法，分析乳癌組織上獨有的脂質與代謝物的特徵型（profiling），做為乳癌亞型診斷之依據。為因應未來更大量的樣品分析與數據處理，本論文利用MATLAB撰寫圖形使用者介面（Graphical User Interface，GUI），將完整的資料前處理的分析流程整合並視覺化，設計一個友善的使用者界面，方便未來所得之新資料進行後續分析。
本論文，描述兩種常壓游離法結合機器學習演算法應用於乳癌診斷的方法開發與其在生物醫學領域、臨床檢驗的適用性。

關鍵字

常壓游離法 ； 紙片電噴灑游離法 ； 脫附電噴灑游離質譜影像 ； 機器學習 ； 小分子代謝物與脂質 ； 癌症診斷

並列摘要

In the past decades, a series of ambient ionization mass spectrometry (AIMS) have been developed and widely introduced to the field of biomedical analysis. In addition, AIMS-based technologies provide a new and effective analysis tool for cancer diagnosis. The advantages of these technologies are that it operates at atmospheric pressure with only minimal sample pretreatment, provides in situ real-time analysis, and rapid mass spectrometric interrogation to the biomedical analysis. Among them, paper spray ionization (PSI) and Desorption Electrospray Ionization (DESI) developed by Electrospray ionization (ESI) based technology have been widely used in medical tests and rapid diagnosis of cancer.
Malignant tumors, also known as cancer, is the leading cause of death in most advanced countries around the world. Breast cancer is the most commonly occurring cancer in women and its incidence rate in Taiwan has reached 70 per 100,000 women. As a result, breast cancer can be seen as one of the diseases that threaten the women’s life. In this thesis, we demonstrated that by combining a lab-built paper spray ionization (PSI) ion source with Orbitrap mass spectrometer to diagnose breast cancer. In routine examinations, core needle biopsy is commonly sampling method in hospitals. Breast tissue is extracted with a hypodermic needle, and wiped onto filter papers for the following pathological diagnosis. We used this routine paper-based tissue sample incorporating with AIMS for a rapid metabolite and lipid profiling to discriminate cancerous/noncancerous tissues. To reduce the chemical noise in the complex matrix without additional sample pretreatment, field asymmetric waveform ion mobility spectrometry was interfaced with paper spray ionization mass spectrometry (FAIMS-PSI-MS). Thus, the predictive metabolite and lipid profiles of core needle biopsied samples could be obtained effectively. By combining of machine learning algorithm and pathological examination, we developed a classification model using small metabolite and lipid profiles for rapid distinction between cancerous and non-cancerous breast tissues.
Furthermore, according the status of hormone receptors like Estrogen receptor (ER) , Progesterone receptor (PR) and human epidermal growth factor receptor 2 (HER2), there are four molecular subtypes of breast cancer including Luminal A, Luminal B, HER2 and Triple negative. The determination of molecular subtypes helps decide the treatment of patients. In this thesis, the DESI-MSI technique was used to directly obtain the chemical information of small metabolite and lipid profiled from breast surgical frozen sections. Each pixel in DESI-MS image was regarded as a data point and was gave the label as "cancer" or "normal" by pathologist using H&E staining. Data points with different molecular subtypes of breast cancer were used as training sets to train the supervised machine learning model by different ML algorithms. Such as naïve Bayes, support vector machine, lasso regression analysis, decision trees, random forests and deep learning method can be used. Our results suggest that combining the data obtained by DESI-MSI with machine learning algorithms can be a new generation approach for diagnosing molecular subtypes in breast cancer by using unique small metabolite and lipid profiles. In order to deal with a larger amount of sample and data processing in the future, we created a Graphical User Interface (GUI) by using MATLAB. This friendly user interface integrate and visualize the overall data pre-processing workflow to facilitate the subsequent analysis of new samples acquired in the future.
In this thesis, the application of two ambient ionization methods combined with machine learning algorithms in breast cancer diagnosis demonstrated that this diagnostic platform has the ability to perform clinical applications.

並列關鍵字

AIMS ； PSI ； DESI ； Machine learning ； small metabolite and lipid ； cancer diagnosis

參考文獻

1. Kaklamanos, G.; Aprea, E.; Theodoridis, G., Mass Spectrometry: Principles and Instrumentation. 2016, 661-668.

2. de Hoffmann, E., Mass Spectrometry. 2005.

3. Finehout, E. J.; Lee, K. H., An introduction to mass spectrometry applications in biological research. Biochem Mol Biol Educ 2004, 32 (2), 93-100.

4. Hsu, C. C.; Dorrestein, P. C., Visualizing life with ambient mass spectrometry. Curr Opin Biotechnol 2015, 31, 24-34.

5. Lebedev, A. T., Ambient ionization mass spectrometry. Russian Chemical Reviews 2015, 84 (7), 665-692.