Nanoparticle–Protein Corona Boosted Cancer Diagnosis with Proteomic Transfer Learning

Keeping pace with the rapid growth of proteomic data,
the integration
of multiproteomic data can improve biomarker identification and cancer
diagnosis. However, the data integration needs to overcome substantial
challenges owing to considerable variability among diverse data set
sources and the extensive range of protein expression levels. In this
study, with serum and urine from the same individuals, we established
two in-depth paired proteome databases, including 956 serum proteins
and 4730 urine proteins. To integrate multiproteomic data, we developed
a proteomic-based transfer learning neural network (ProteoTransNet)
to enhance the accuracy of bladder cancer diagnosis and progression
monitoring. Using random forest analysis on the integrated database,
we selected two panels comprising the top 10 key proteins, achieving
a diagnostic AUC of 0.996 and a stage classification AUC of 0.914.
ProteoTransNet integrates serum and urine proteome databases with
proteomic transfer learning, significantly enhancing the diagnostic
accuracy through minimizing biases and errors caused by variations
in proteomic data. Our study provides insights that transfer learning
of sophisticated biological information may solve complicated biological
problems in disease diagnosis, prognosis, and treatment.

伴随蛋白质组学数据的快速增长，多蛋白质组数据整合可有效提升生物标志物识别与癌症诊断效能。然而，由于不同数据集来源间存在显著异质性，且蛋白质表达水平跨度极广，数据整合仍需攻克诸多严峻挑战。本研究选取同一受试者的血清与尿液样本，构建了两套深度覆盖的配对蛋白质组数据库，涵盖956种血清蛋白与4730种尿液蛋白。为实现多蛋白质组数据整合，本研究开发了一种基于蛋白质组学的迁移学习神经网络（ProteoTransNet），以提升膀胱癌诊断与进展监测的准确性。通过对整合数据库开展随机森林分析，本研究筛选出两组各包含10种关键蛋白的蛋白特征面板，其诊断受试者工作特征曲线下面积（AUC）达0.996，分期分类AUC达0.914。ProteoTransNet通过蛋白质组学迁移学习整合血清与尿液蛋白质组数据库，通过最小化由蛋白质组数据异质性引发的偏倚与误差，显著提升了诊断准确性。本研究揭示，对复杂生物信息开展迁移学习，或可解决疾病诊断、预后评估与治疗中的复杂生物学难题。