用機器學習方法尋找生物標記以預測PD-L1抑製劑於治療癌症的臨床反應

Abstract

免疫治療的其中一種方法是用藥物阻斷PD-1/PD-L1的蛋白質交互作用，激活免疫系統，使其攻擊腫瘤細胞。然而，該療法效於超過五成的病人無效。如果我們能夠找到導致藥物無效的生物標記基因，就可以辨認出適合使用免疫療法的病患。 我們發展了一個使用基因表現量作為變量的模型，用於預測癌症病人接受Atezolizumab (一種PD-L1 抑製劑)藥物治療後的臨床反應。我們用的分類方法是logistic ridge regression。並使用DESeq2、BSS-WSS ratio、及一種領域自適應的方法選取基因。訓練資料集是泌尿上皮癌(mUC)患者。該模型在另一組mUC病人的測試資料集中的預測AUC為0.76。同一組的mUC病人訓練資料集亦用於建立非小細胞肺癌(NSCLC)和腎癌(RCC)患者適用的分類器，並分別預測AUC為 0.69及0.70。這項研究發現了一些生物標記基因，包括 CXCL9、LURAP1、LYRM1、SIGLEC17P 和 UST。此外，我們比較不同的分類器用於單一癌症資料集的表現，發現線性模型(包括logistic ridge和SVM-linear)優於非線性模型(包括SVM-RBF、AdaBoost、CatBoost和XGBoost)。 再者，我們研究那一種分類方法能預測mUC、NSCLC和RCC三種癌症的病人（一組跨癌症的資料集），並與Banchereau et al. (2021) 發表的研究比較。該研究的分類方法是logistic LASSO regression，預測AUC是0.62。我們發現XGBoost的AUC是0.64，使用提升法(boosting)可以提高模型的表現。 它優於一個使用logistic ridge、SVM-linear、SVM-RBF、AdaBoost、CatBoost和XGBoost的堆疊法(stacking)分類器，其預測AUC為0.61。

Immunotherapy by PD-1/PD-L1 blockade induces durable clinical responses in cancer patients. However, a portion of patients are resistant to this medication treatment. If we could find the biomarker genes causing the failure, we could stratify patients who would respond to the therapy. We proposed a gene-based classifier which takes gene expression as input for clinical response prediction to a PD-L1 inhibitor, Atezolizumab. It is a logistic ridge regression classifier using genes identified by DESeq2, ranked by the BSS-WSS ratio, and filtered by a domain adaptation technique. The training set is a dataset of metastasis urothelial cancer (mUC) patients. The classifier has an AUC of 0.76 in a test set of mUC patients. Two other classifiers were developed using the same training set of mUC patients, and have test sets of non-small cell lung cancer (NSCLC) or renal cell cancer (RCC) patients. The models have AUCs of 0.69 and 0.70, respectively. Some biomarkers, including CXCL9, LURAP1, LYRM1, SIGLEC17P, and UST, have been discovered. Moreover, multiple machine learning methods are examined in this study. We have found that the linear models (such as logistic ridge and SVM-linear) outperform the non-linear models (such as SVM-RBF, AdaBoost, CatBoost, and XGBoost) in cancer-specific datasets. In addition, we investigated classification methods using a cross-cancer dataset of mUC, NSCLC, and RCC patients, and compared them to that in Banchereau et al. (2021), which is a logistic LASSO classifier with an AUC of 0.62. We found that XGBoost has an AUC of 0.64. The boosting technique can improve the model's performance. It is superior to a stacking classifier that aggregates logistic ridge, SVM-linear, SVM-RBF, AdaBoost, CatBoost, and XGBoost, which has an AUC of 0.61.