Wang Xiaosong, Feng Haisong, Zhang Yilei, Lin Fan
School of Information and Artificial Intelligence, Anhui Agricultural University, Hefei, Anhui, 230036, China.
School of Informatics, Xiamen University, Xiamen, Fujian, 361105, China.
Comput Biol Med. 2025 May;190:110067. doi: 10.1016/j.compbiomed.2025.110067. Epub 2025 Mar 26.
Somatic mutations could influence critical cellular processes, leading to uncontrolled cell growth and tumor formation. Understanding the intricate interactions between somatic mutations and drugs was crucial for advancing our knowledge of the underlying biological mechanisms of cancer. This knowledge, in turn, could drive advancements in cancer detection, diagnosis, and treatment. Exploring the relationships between specific somatic mutations and drug responses held the potential to identify targeted therapeutic interventions and improve personalized treatment strategies for cancer patients. In this study, we introduced a computational model, the signed graph comparison learning for mutation-drug associations (SGCLMD), designed to predict signs of somatic mutation-drug associations. Initially, we leveraged clinical data to construct a benchmark dataset encompassing somatic mutation-drug associations. We proposed a graph enhancement method, employing a random perturbation strategy, to expand the signed graph. This approach not only preserved interaction information across the two perspectives of the signed graph but also retained implicit relationships between these perspectives. Furthermore, we devised a multi-view comparison loss algorithm to learn node representations for the graph generated post-random perturbation. Through parameter optimization using 5-fold cross-validation, our SGCLMD model achieves optimal area under the curve (AUC) and area under the precision-recall curve (AUPR) values of 0.8306 and 0.8751, respectively, representing improvements of 3 % and 3.1 % over the state-of-the-art method. Through ablation experiments and case studies, we validated the importance of graph enhancement methods and multi-view contrast learning modules, demonstrating SGCLMD's potential in predicting somatic mutation-drug associations. The code and dataset for SGCLMD are available at https://github.com/wangxiaosong96/SGCLMD.
体细胞突变可能影响关键的细胞过程，导致细胞不受控制地生长和肿瘤形成。了解体细胞突变与药物之间复杂的相互作用对于推进我们对癌症潜在生物学机制的认识至关重要。反过来，这些知识可以推动癌症检测、诊断和治疗的进步。探索特定体细胞突变与药物反应之间的关系有可能识别出靶向治疗干预措施，并改善癌症患者的个性化治疗策略。在本研究中，我们引入了一种计算模型，即用于突变 - 药物关联的带符号图比较学习（SGCLMD），旨在预测体细胞突变 - 药物关联的正负性。最初，我们利用临床数据构建了一个包含体细胞突变 - 药物关联的基准数据集。我们提出了一种图增强方法，采用随机扰动策略来扩展带符号图。这种方法不仅保留了带符号图两个视角的交互信息，还保留了这些视角之间的隐含关系。此外，我们设计了一种多视图比较损失算法，用于学习随机扰动后生成的图的节点表示。通过使用五折交叉验证进行参数优化，我们的SGCLMD模型分别实现了0.8306和0.8751的最优曲线下面积（AUC）和精确召回曲线下面积（AUPR）值，比现有方法分别提高了3%和3.1%。通过消融实验和案例研究，我们验证了图增强方法和多视图对比学习模块的重要性，证明了SGCLMD在预测体细胞突变 - 药物关联方面的潜力。SGCLMD的代码和数据集可在https://github.com/wangxiaosong96/SGCLMD获取。
