Brlek Petar, Bulić Luka, Shah Nidhi, Shah Parth, Primorac Dragan
St. Catherine Specialty Hospital, 10000 Zagreb, Croatia.
School of Medicine, Josip Juraj Strossmayer University of Osijek, 31000 Osijek, Croatia.
Int J Mol Sci. 2025 Mar 13;26(6):2568. doi: 10.3390/ijms26062568.
Cancers of unknown primary (CUPs) represent a significant diagnostic and therapeutic challenge in the field of oncology. Due to the limitations of current diagnostic tools in these cases, novel approaches must be brought forward to improve treatment outcomes for these patients. The objective of this study was to develop a machine-learning-based software for primary cancer site prediction (OncoOrigin), based on genetic data acquired from tumor DNA sequencing. By design, this was an diagnostic study, conducted using data from the cBioPortal database (accessed on 21 September 2024) and several data processing and machine learning Python libraries. The study involved over 20,000 tumor samples with information on patient age, sex, and the presence of genetic variants in over 600 genes. The main outcome of interest was machine-learning-based discrimination between cancer site classes. Model quality was assessed by training set cross-validation and evaluation on a segregated test set. Finally, the optimal model was incorporated with a graphical user interface into the OncoOrigin software. Feature importance for class discrimination was also determined on the optimal model. Out of the four tested machine learning estimators, the XGBoostClassifier-based model proved superior in test set evaluation, with a top-2 accuracy of 0.91 and ROC-AUC of 0.97. Unlike other machine learning models published in the literature, OncoOrigin stands out as the only one integrated with a graphical user interface, which is crucial for facilitating its use by oncology specialists in everyday clinical practice, where its application and implementation will have the greatest value in the future.
原发灶不明的癌症（CUPs）是肿瘤学领域一项重大的诊断和治疗挑战。由于当前诊断工具在这些病例中的局限性，必须提出新的方法来改善这些患者的治疗效果。本研究的目的是基于从肿瘤DNA测序获得的基因数据，开发一种基于机器学习的原发性癌症部位预测软件（OncoOrigin）。从设计上来说，这是一项诊断性研究，使用了来自cBioPortal数据库（于2024年9月21日访问）的数据以及几个数据处理和机器学习的Python库。该研究涉及超过20000个肿瘤样本，包含患者年龄、性别以及600多个基因中基因变异的存在情况等信息。主要关注的结果是基于机器学习对癌症部位类别进行区分。通过训练集交叉验证和在隔离测试集上的评估来评估模型质量。最后，将最优模型与图形用户界面整合到OncoOrigin软件中。还在最优模型上确定了类别区分的特征重要性。在四个测试的机器学习估计器中，基于XGBoostClassifier的模型在测试集评估中表现出色，前两名准确率为0.91，ROC-AUC为0.97。与文献中发表的其他机器学习模型不同，OncoOrigin是唯一集成了图形用户界面的模型，这对于肿瘤学专家在日常临床实践中使用它至关重要，其应用和实施在未来将具有最大价值。
