整合核糖核酸測序與臨床數據與使用具有子圖表式的圖模型預測癌症預後

Abstract

疾病是人類的主要死因，其中又以癌症占最大比例。這表明，通過深入的癌症研究，我們能夠對全球死亡率的降低作出實質性的貢獻。癌症的預後預測是一個重要議題。早期階段的高風險患者經過治療後可以達到很高的存活率。透過數據科學與深度學習的幫助，我們能從患者身上得到不同性質資料的隱藏資訊。高通量 (High-throughput) 技術就是其中的一種，它可以產生出大量組學數據供我們分析。預後預測中使用到的基因組學數據便是其中一種，然而它具備樣本少、特徵多兩種特性，在這種情況下，模型難以捕捉到重要訊息。為了解決這個狀況，我們使用先前開發的系統生物學特徵選擇器 (Systems Biology Feature Selector) 進行降維，它會從 RNA 測序 (RNA-Seq) 數據中選擇與癌症預後高度相關的生物標誌物 (biomarker)，並且引入了基因相互作用網路 (Gene Interaction Networks) 來增加額外的信息。為了充分使用這兩種資料，我們選用圖神經網路 (Graph Neural Network; GNN) 作為我們的基底模型。將篩選出來的預後生物標誌物視為節點，基因相互作用網路中獲得的基因之間的潛在關係視為邊。考慮到不同的基因組合可能含有隱藏的資訊，我們將預後生物標誌物構成的圖拆分為更小的子圖，並利用獨立的圖卷積層萃取子圖的資訊，此子圖集的圖捲積宛如卷積神經網路 (Convolutional Neural Network; CNN) 中的感受視野 (receptive field)，試圖抓到大圖中的圖案 (pattern)。此外，我們也納入了臨床數據，設計了子圖級圖雙模態神經網路，它不但能抓取不同基因子集包含的隱藏關係，也能從RNA測序與臨床數據中提取有用資訊。實驗結果顯示出良好的性能，特別是在精確召回曲線下面積 (Area Under the Precision-Recall Curve; AUPRC) 中，與先前和基線模型相比平均提高了 9.16% 和 9.8% ，這間接證明了基因子集合擁有豐富資訊。我們希望此研究能為癌症研究盡一份心力。

Diseases are a leading cause of human mortality, with cancer comprising a significant proportion. Conducting robust research on cancer can contribute to a global reduction in mortality rates. Prognostic prediction for cancer is a pressing issue, as high-risk patients in the early stages of the disease can achieve high survival rates through timely intervention. Leveraging data science and deep learning enables us to uncover hidden insights from diverse patient data. High-throughput technologies, such as genomics, generate extensive datasets for analysis. However, genomic data used in prognosis prediction presents limited samples and high-dimensional features, making it challenging for models to capture crucial information in such extreme conditions. To address this challenge, we employ a previously developed systems biology feature selector for dimensionality reduction. This selector identifies biologically relevant biomarkers highly correlated with cancer prognosis from RNA sequencing data. Additionally, we incorporate gene interaction networks to add additional information. To fully harness these two types of data, we opt for a graph neural network as our foundational model. We treat the selected prognostic biomarkers as nodes and the potential relationships between genes obtained from the gene interaction network as edges. Considering that different gene combinations may contain hidden information, we decompose the graph formed by prognostic biomarkers into smaller subgraphs. Using independent graph convolutional layers, we extract special feature embeddings between gene subsets from the subgraph. This subgraph-level graph convolution mimics the receptive field in convolutional neural networks, aiming to capture patterns in the larger graph. Furthermore, we integrate clinical data and design a subgraph-level graph bimodal neural network. This network not only captures hidden relationships within different gene subsets but also extracts valuable information from RNA sequencing and clinical data. Our experimental results demonstrate promising performance, particularly for the AUPRC, exhibiting an average improvement of 9.16% and 9.8% when compared to our previous approach and the baseline models. This indirectly verifies the substantial information richness within the gene subset. With this study, we aim to drive meaningful progress in the field of cancer research.