| 摘要: | 機器學習作為一項強大高效的工具，已經被廣泛應用在多個專業領域上，在大數據時代中顯著提高了各項任務的準確性以及執行效率。而在擁有大量樣本需要處理的生物醫學中，更是擁有著巨大潛力。本研究聚焦於使用基於機器學習的一系列特徵選擇方法，探究個體基因位點變異與乳癌治療藥物的四種副作用(骨質疏鬆、周邊神經病變、白血球數量異常與子宮內膜厚度)的關係。與探索特定副作用與基因位點之間關係的主流研究不同，我們的研究著重於揭示藥物副作用與基因變異位點之間的關聯性，希望確保患者的安全並選擇更為恰當的治療方法。
通過多階段分析，我們篩選出了與四種副作用各自高度相關的基因變異位點。在統計學上，我們比較了多種特徵選擇方法(卡方檢驗、費雪精確測試、斯皮爾曼等級相關係數、肯德爾相關係數，並使用p < 0.05作為閾值)進行第一階段的特徵選擇，接著評估了不同的機器學習分類器(隨機森林、XGBoost、深度神經網路)的差異作為第二階段的特徵選擇器。接著對各個特徵選擇器的分類準確率以及學習曲線進行深入比較，探討特徵重要性及基因位點對預測副作用有無的影響。
我們的研究針對每個副作用，將超過150,000個獨立的基因位點篩選至約100個關鍵位點，顯著提高了藥物副作用的預測準確性。此外，為了取得更詳細的驗證，我們針對與藥物代謝與免疫密切相關的HLA基因型進行分型分析，接著和機器學習模型篩選出的重要基因位點互相比較，從而提供了對藥物副作用機制更全面的理解。
;Machine learning is a powerful and efficient tool that has been widely applied across various professional fields, significantly enhancing the accuracy and execution efficiency of tasks in the era of big data. In the biomedical field, where large sample sizes need to be processed, machine learning demonstrates substantial potential. This study focuses on employing various machine learning-based feature selection methods to investigate the relation between individual gene locus variations and four side effects of breast cancer treatment: Osteoporosis, Peripheral Neuropathy, abnormal Endometrial Thickness and White Blood Cell Count. Unlike mainstream research that explores the relation between specific illnesses and genetic loci, our research focuses on unveiling the association between treatment side effects and genetic loci to ensure patient safety and select appropriate treatment methods.
Through multi-stage analysis, we identified genetic variant loci highly correlated with each of the four types of side effects. We compared various feature selection methods (Chi-Square, Fisher exact, Spearman’s Rank, Kendall Tou, using p value < 0.05 as the threshold) for the first stage of feature selection. Subsequently, we evaluated different machine learning classifiers (Random Forest, XGBoost, Neural Network) as the second stage of feature selectors. We conducted in-depth comparisons of the accuracy and learning curves of each feature selector, analyzing the importance of features and the impact of genetic loci on predicting side effects.
Our research narrowed down over 150,000 independent genetic loci to approximately 100 key loci for each side effect, significantly improving the accuracy of predicting medication side effects. To obtain further validation, we conducted genotype analysis on HLA, which is closely related to drug metabolism and immunity. |
