整合多體學與可解釋人工智慧應用於醣基化之癌症臨床研究及個人醫療

Abstract

在腫瘤生成與轉移的各個階段皆有醣基化與腫瘤之間的交互作用，並且其臨床潛力亦有大量證據加以證實。藉由次世代定序的腫瘤數據及多體學的整合分析，我們了解到此關係不僅僅是展現於全基因體或蛋白體，醣基化在腫瘤間也有不同之處。隨著剖析醣基化與腫瘤關係的需求增長，我們系統性地識別出三千多個醣基化相關基因 (醣基因)，並整合基因體與表關基因體證明其在多癌症分類的臨床潛力。藉由將高維度的多體學資料進行低維度映射後，更進一步彰顯醣基因以及醣基化相關生物路徑皆具有揭示與組織起源相關的可能癌症簇的能力，並且與全基因組的結果相當。另外不同癌症簇亦呈現相異的分子概況、免疫特徵、細胞週期活動和存活分佈。最後我們進一步通過機器學習（XGBoost 和隨機生存森林）和 XAI（可解釋人工智能）的進階整合，描繪跨越各種癌症類型的潛在診斷和預後標誌物，為每個癌症簇構建了五年生存預測模型。更重要的是，我們亦使用此方法建立個人化的生存預測模型。盼能以此醣基化為基底的機器學習模型用於個人化醫療與治療方針等臨床應用。

Crosstalk between glycosylation and tumor already poses its clinical potential regarding the comprehensive involvement during tumorigenesis and downsteam metastasis. Also, bulk sequencing and the cutting-edged approaches are manifesting the relationship through omic integration instead of transcriptome or proteome alone but the framework of such glycosylation differs. With the emerging need of deciphering the interaction insightfully and comprehensively, we systematically identified over 3,000 glycogenes and subsequent omic integration further certified their potential in multi-cancer classification through low-dimension mapping. Such mapping also revealed the latent embedding of glycome, either in gene or pathway level. Similar with whole genome-based profile, this specialized geneset also disclose possible cancer clusters correlated to tissue-origin whilst presenting distinct molecular profile, immune signature, cell cycle activities, and survival distributions. With the advanced integration of machine learning (XGBoost and Random Survival Forest) and XAI (eXplainable artificial intelligence), we depicted the potential diagnostic and prognostic markers crossing various cancer types, constructed five-year survival prediction models for each cancer cluster, and also had individualized survival predictions for further clinical implementations including personalized medicine and treatment guideline.