生物多模态数据可以提供基因调控和细胞发育的全面视图，对于复杂疾病的研究具有重要意义。癌症对人民的健康造成了极大威胁。然而当今癌症诊疗领域面临着缺乏高敏感度、合理可量化的筛查和诊疗方案等共性瓶颈问题。近年来，生物多模态数据的研究在癌症诊疗领域的重要性逐渐显现，但对于如何实现多模态数据的有效整合，以及如何建立有效的癌症诊疗体系，还有待探索和研究。 本论文首先提出了针对癌症无创诊断的cfRNA多模态标志物整合的机器学习模型。本论文开发了结合文献系统性审查、统计分析和实验验证的肝癌多模态标志物筛选流程，筛选得到1个lncRNA、1个miRNA、3个cfRNA片段和1个cfRNA可变剪接标志物。最终，本论文利用机器学习组合6个cfRNA多模态标志物和蛋白质标志物AFP，在患有肝炎肝硬化的高危人群中进行早期肝癌诊断，实验结果中模型AUC达到0.922，表现出良好的敏感性（84%）和特异性（86%）。 传统的生物多模态整合方法忽视了先验的生物知识，通用性差且缺乏可解释性。因此，本论文创新性地提出基于生物通路启发的深度学习模型（Pathformer模型），首次将先验的生物通路串扰信息引入Transformer模型，并应用于生物多模态数据整合。Pathformer利用压缩的多模态向量和基于通路的稀疏神经网络处理多模态输入，利用十字交叉注意力机制来捕捉不同生物通路和模态之间的串扰，并利用Shapley加法解释方法来揭示关键调节机制，在一定程度上缓解了 “黑箱”问题。通过在32个基准数据集的测试，本论文证明了Pathformer优于其他18个模型，在癌症预后预测中平均F1得分提高6.3%-14.7%；在癌症分期预测中平均F1分数提高5.1%-12%；在癌症分型预测中平均F1得分提高1.7%-12.7%；在癌症药物反应预测中平均F1分数提高8.1%-13.6%。通过消融分析，本论文证明了不同模态和Pathformer的不同模块对于生物多模态整合的必要性。 最后，本论文将Pathformer应用到癌症组织和血液的多模态数据中，以解决癌症无创诊断、分型、分期、预后和药物反应预测等癌症诊疗问题，并通过模型的生物可解释模块发现了许多与疾病相关的潜在调控机制。 综上所述，本论文围绕着“生物多模态数据整合”，构造了用于癌症无创诊断多模态标志物整合的机器学习模型，创新性地开发了基于生物通路启发的深度学习模型，能够缓解深度学习模型的不可解释性，并解决了不同的癌症诊疗问题。
Multi-modal biological data can offer a comprehensive view of gene regulation and cellular development from various levels and perspectives. Therefore, multi-modal biological data integration is crucial for research into various complex diseases. Cancer poses a significant threat to the health and well-being of people. The lack of highly sensitive early cancer screening methods and quantifiable treatment strategies is a common bottleneck in the field of cancer diagnosis and therapy. In recent years, the importance of utilizing multi-modal biological data for cancer diagnosis and treatment has become increasingly evident. However, the effective integration of multi-modal biological data and the establishment of an efficient cancer diagnostic and therapeutic system remain to be explored and researched.We initially attempted to utilize machine learning to integrate multi-modal features of cell-free RNA (cfRNA) for cancer non-invasive diagnosis. We developed a hepatocellular carcinoma (HCC) multi-modal biomarker selection process that combines systematic literature review, statistical analysis, and experimental validation. This process identified multi-modal biomarkers including 1 lncRNA, 1 miRNA, 3 cfRNA fragments, and 1 cfRNA alternative splicing biomarker. Ultimately, we constructed a machine learning model to combine these six cfRNA multi-modal biomarkers and the protein biomarker AFP. The machine learning model achieved an AUC of 0.922, along with a sensitivity of 84% with specificity of 86%, for early hepatocellular carcinoma detection in high-risk populations with hepatitis and liver cirrhosis.Conventional integration methods ignore prior biological knowledge, suffering from poor generalizability and a lack of interpretability. To address this limitation, we introduced an innovative deep learning model based on biological pathway informed (Pathformer) for multi-modal biological data integration, which incorporates prior biological pathway crosstalk information as bias into a Transformer-based framework for the first time. Pathformer embeds multi-omics input with a compacted multi-modal vector and a pathway-based sparse neural network, leverages criss-cross attention mechanism to capture crosstalk between different biological pathways and modalities, and utilizes Shapley Additive Explanation method to reveal key pathways, genes, and regulatory mechanisms, partially alleviating the “black-box” issue of AI models. Through testing on 32 benchmark datasets, we demonstrated that Pathformer outperformed 18 other methods, with an average improvement in F1 score of 6.3%-14.7% for cancer survival prediction, 5.1%-12% for cancer stage prediction, 1.7%-12.7% for cancer typing prediction, and 8.1%-13.6% for cancer drug response prediction. Through ablation analysis, we substantiated the necessity of different modalities and Pathformer‘s distinct modules for biological multi-modal data integration.Finally, we systematically applied Pathformer to various multi-modal data from tissue and blood, addressing a range of cancer diagnosis and treatment issues including non-invasive diagnosis, subtyping, staging, prognosis, and drug response prediction. We have identified many interesting disease-related deregulation events using Pathformer.In summary, we conducted research around the theme of " biological multi-modal data integration". We constructed a machine learning model for integrating multi-modal biomarkers in non-invasive cancer diagnosis, innovatively developed a deep learning model inspired by biological pathways information, partially alleviated the “black-box” issue of AI models, and applied it to address various cancer diagnosis and treatment challenges.
