淋巴結轉移的預測在腫瘤學中仍然是一項重大挑戰，尤其是在小樣本資料集中，傳統預測方法往往難以捕捉到生物意義。本研究開發了一個神經網路模型，整合基因表達與單細胞的細胞間通訊，用於預測淋巴結轉移。此模型是一個具有生物學知識導向的神經網路，設計目標是具有可解釋性的同時保有一定的預測準確率。我們透過交叉驗證評估模型表現，並與傳統機器學習方法進行比較，在TCGA測試集中，此模型達成0.771的AUROC；在兩組獨立測試集GSE43580和GSE50081中，分別達到0.664與0.771，表現與現有研究成果相當。除了預測準確率外，我們也利用深度學習的可解釋性，找出潛在的生物標記與訊號通路。這證實我們的方法在預測效能與生物意義解釋性之間取得了良好平衡。本研究顯示深度學習在提升轉移預測模型方面的潛力，並為未來將生物交互資訊整合進深度學習模型奠定基礎。;Lymph node metastasis prediction remains a critical challenge in oncology, especially for small datasets where traditional methods struggle to extract meaningful patterns. This study develops a neural network model that integrates gene expression data and cell-cell communication features to predict lymph node metastasis. Our model is a biologically informed neural network, specifically designed to be interpretable while preserving high predictive accuracy. We evaluate its performance through cross-validation and compare it against traditional machine learning methods. Our model achieved an area under receiver operating characteristic curve (AUROC) of 0.771 on The Cancer Genome Atlas program (TCGA) testing set, and 0.664 and 0.771 on independent test sets GSE43580 and GSE50081, respectively, which is comparable to current research results. In addition to predictive accuracy, our model leverages interpretability from deep learning to identify potential biomarkers and signaling pathway. This demonstrates that our approach achieves a balance between predictive performance and biological interpretability. This study highlights the potential of deep learning in improving predictive modeling for metastasis risk assessment and provides a foundation for future research integrating biological interactions into computational models.
