预测耐药性与伪进展：极简免疫编辑数学模型是否能预测肺癌中检查点抑制剂治疗的结局？

Predicting resistance and pseudoprogression: are minimalistic immunoediting mathematical models capable of forecasting checkpoint inhibitor treatment outcomes in lung cancer?

DOI 原文链接

用sci-hub下载

MATHEMATICAL BIOSCIENCES

影响因子:1.8

分区:数学4区 / 生物学4区数学与计算生物学4区

发表日期:2024 Oct

作者: Kevin Robert Scibilia, Pirmin Schlicke, Folker Schneller, Christina Kuttler

DOI: 10.1016/j.mbs.2024.109287

摘要

免疫检查点抑制剂（ICIs）靶向PD-1/PD-L1在肺癌治疗中的应用日益增加，迫切需要可靠预测个体患者的治疗效果。为弥合预测差距，我们研究了四种不同的数学模型（包括一种新颖的延迟反应模型），采用常微分方程形式。我们严格评估它们基于模型结果概率的个体及组合预测能力，重点关注患者的进行性疾病（PD）状态。通过拟合完整治疗过程，发现新型延迟反应模型（R2=0.938）优于最简模型（R2=0.865）。模型组合能够通过仅用原发肿瘤最大直径测量值，可靠预测患者PD结局，整体准确率达77%（敏感性=70%，特异性=81%），且能自主识别51%的患者子集，其预测整体准确率为81%（敏感性=81%，特异性=81%）。所有模型均显著优于完全数据驱动的机器学习方法。这些建模方法提供了一个动态的基础框架，支持临床医生通过已可用的临床测量数据，识别不同的治疗结果轨迹。结合其他预测工具和生物标志物的联合应用，以及更多疾病信息（如转移阶段），可进一步提升治疗结局的预测能力。我们认为，简单的模型形式有助于在其他癌症类型中广泛应用，也可以为其他治疗方式设计类似模型。

Abstract

The increased application of immune checkpoint inhibitors (ICIs) targeting PD-1/PD-L1 in lung cancer treatment generates clinical need to reliably predict individual patients' treatment outcomes.To bridge the prediction gap, we examine four different mathematical models in the form of ordinary differential equations, including a novel delayed response model. We rigorously evaluate their individual and combined predictive capabilities with regard to the patients' progressive disease (PD) status through equal weighting of model-derived outcome probabilities.Fitting the complete treatment course, the novel delayed response model (R2=0.938) outperformed the simplest model (R2=0.865). The model combination was able to reliably predict patient PD outcome with an overall accuracy of 77% (sensitivity = 70%, specificity = 81%), solely through calibration with primary tumor longest diameter measurements. It autonomously identified a subset of 51% of patients where predictions with an overall accuracy of 81% (sensitivity = 81%, specificity = 81%) can be achieved. All models significantly outperformed a fully data-driven machine learning-based approach.These modeling approaches provide a dynamic baseline framework to support clinicians in treatment decisions by identifying different treatment outcome trajectories with already clinically available measurement data.Conjoint application of the presented approach with other predictive tools and biomarkers, as well as further disease information (e.g. metastatic stage), could further enhance treatment outcome prediction. We believe the simple model formulations allow widespread adoption of the developed models to other cancer types. Similar models can easily be formulated for other treatment modalities.