PO.CL01.18 · 临床研究
优化的甲基化-蛋白质多癌种早期检测（MCED）检测分类器的性能
Performance of an optimized methylation-protein multi-cancer early detection (MCED) test classifier
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：多癌种早期检测（MCED）检测可检测多种癌症类型和分期。我们此前开发了一种甲基化和蛋白质（MP V1）MCED分类器，目标特异性≥98.0%（Gainullin，medRxiv，2025.08.24.25334244）。在此，我们描述一种经过改进的MP V2分类器，其基于对提升性能的分类器模型架构的评估而确定。
方法：与MP V1相比，MP V2分类器经过训练，在病例-对照目标特异性≥97.0%下实现提高的早期敏感性。MP V2分类器架构使用一个训练集（654例癌症；2,373例非癌症）开发，该训练集被划分为5折交叉验证和小型留出集。使用一个独立的小型留出测试集（110例癌症；509例非癌症）比较锁定的候选模型。使用一个先前描述的测试集（729例癌症；2,434例非癌症）比较MP V1和MP V2分类器的性能，MP V2性能还在一个设计得更贴近预期使用人群的独立临床验证测试集（324例癌症；800例非癌症）中进行了评估。
结果：与MP V1相比，MP V2（特异性97.4%）在测试集中显示总体敏感性提高7.3%，I期、II期、I/II期、III期、IV期和未知分期的敏感性分别提高7.6%、9.2%、8.3%、8.0%、3.8%和13.3%（表1）。在独立验证集中，MP V2（特异性97.4%；95% CI：96.0-98.3%）总体敏感性为41.4%，I期、II期、I/II期、III期和IV期的敏感性分别为16.0%、31.3%、22.8%、52.9%和83.1%。排除乳腺癌和前列腺癌后，验证集中MP V2总体敏感性为55.6%，I期、II期、I/II期、III期和IV期的敏感性分别为26.8%、42.9%、34.8%、63.6%和89.3%。
结论：在病例-对照环境中，MP V2分类器在较低的特异性目标下为早期癌症提供了改善的敏感性。
表1. MP V1和MP V2分类器在测试集中的性能。a排除乳腺癌和前列腺癌。MP V1分类器性能（95% CI）MP V2分类器性能（95% CI）测得特异性（N=2,434）98.5%（97.9-98.9）97.4%（96.7-97.9）总体敏感性，所有癌症（N=729）50.9%（47.3-54.5）57.8%（54.1-61.3）I期（n=182）15.4%（10.9-21.3）21.4%（16.1-27.9）II期（n=163）38.0%（30.9-45.7）46.0%（38.5-53.7）III期（n=180）67.8%（60.6-74.2）76.1%（69.4-81.8）IV期（n=172）85.5%（79.4-90.0）89.5%（84.1-93.3）未知分期（n=32）37.5%（22.9-54.7）50.0%（33.6-66.4）I/II期（n=345）26.1%（21.7-31.0）33.0%（28.3-38.2）总体敏感性（N=590）a 56.8%（52.8-60.7）64.1%（60.1-67.8）I期（n=145）17.2%（12.0-24.2）24.8%（18.5-32.4）II期（n=109）48.6%（39.4-57.9）57.8%（48.4-66.6）III期（n=151）73.5%（66.0-79.9）81.5%（74.5-86.8）IV期（n=155）86.5%（80.2-91.0）90.3%（84.6-94.0）未知分期（n=30）40.0%（24.6-57.7）53.3%（36.1-69.8）I/II期（n=254）30.7%（25.4-36.6）39.0%（33.2-45.1）
查看英文原文 English abstract
Background: Multi-cancer early detection (MCED) tests can detect several cancer types and stages. We previously developed a methylation and protein (MP V1) MCED classifier with a target specificity of ≥98.0% (Gainullin, medRxiv, 2025.08.24.25334244). Herein, we describe a refined MP V2 classifier that was identified based on the evaluation of classifier model architectures that improved performance.
Methods: Compared to MP V1, the MP V2 classifier was trained to achieve increased early-stage sensitivity at a case-control target specificity of ≥97.0%. MP V2 classifier architecture was developed using a training set (654 cancer; 2,373 non-cancer) partitioned into 5-fold cross validation and mini-holdout sets. Locked candidate models were compared using an independent mini-holdout test set (110 cancer; 509 non-cancer). MP V1 and MP V2 classifier performance were compared using a previously described test set (729 cancer; 2,434 non-cancer), and MP V2 performance was also evaluated in an independent clinical validation test set (324 cancer; 800 non-cancer) that was designed to more closely mimic the intended use population.
Results: Compared to MP V1, MP V2 (specificity of 97.4%) demonstrated a 7.3% increase in overall sensitivity, with sensitivity increases of 7.6%, 9.2%, 8.3%, 8.0%, 3.8%, and 13.3% for stages I, II, stages I/II, III, IV, and unknown, respectively, in the test set (Table 1). In the independent validation set, MP V2 (specificity 97.4%; 95% CI: 96.0-98.3%) overall sensitivity was 41.4%, with sensitivities of 16.0%, 31.3%, 22.8%, 52.9%, and 83.1% for stages I, II, stages I/II, III and IV, respectively. Excluding breast and prostate cancers, MP V2 overall sensitivity was 55.6%, with sensitivities of 26.8%, 42.9%, 34.8%, 63.6%, and 89.3% for stages I, II, stages I/II, III, and IV, respectively, in the validation set.
Conclusion: In a case-control setting, the MP V2 classifier offered improved sensitivity for early-stage cancers at a lower specificity target.
Table 1. MP V1 and MP V2 classifier performance in the test set. a breast and prostate excluded. MP V1 Classifier Performance (95% CI) MP V2 Classifier Performance (95% CI) Measured Specificity (N=2,434) 98.5% (97.9-98.9) 97.4% (96.7-97.9) Overall Sensitivity, all cancers (N=729) 50.9% (47.3-54.5) 57.8% (54.1-61.3) Stage I (n=182) 15.4% (10.9-21.3) 21.4% (16.1-27.9) Stage II (n=163) 38.0% (30.9-45.7) 46.0% (38.5-53.7) Stage III (n=180) 67.8% (60.6-74.2) 76.1% (69.4-81.8) Stage IV (n=172) 85.5% (79.4-90.0) 89.5% (84.1-93.3) Unknown Stage (n=32) 37.5% (22.9-54.7) 50.0% (33.6-66.4) Stages I/II (n=345) 26.1% (21.7-31.0) 33.0% (28.3-38.2) Overall Sensitivity, (N=590) a 56.8% (52.8-60.7) 64.1% (60.1-67.8) Stage I (n=145) 17.2% (12.0-24.2) 24.8% (18.5-32.4) Stage II (n=109) 48.6% (39.4-57.9) 57.8% (48.4-66.6) Stage III (n=151) 73.5% (66.0-79.9) 81.5% (74.5-86.8) Stage IV (n=155) 86.5% (80.2-91.0) 90.3% (84.6-94.0) Unknown Stage (n=30) 40.0% (24.6-57.7) 53.3% (36.1-69.8) Stages I/II (n=254) 30.7% (25.4-36.6) 39.0% (33.2-45.1)
利益披露 Disclosure
V. G. Gainullin,
Exact Sciences Corp. Employment, Stock.
M. Gray,
Exact Sciences Corp. Employment, Stock.
M. Kumar,
Exact Sciences Corp. Employment, Stock.
S. Luebker,
Exact Sciences Corp. Employment, Stock.
A. Lehman,
Exact Sciences Corp. Employment, Stock.
D. D. Flake,
Exact Sciences Corp. Employment, Stock.
A. Shanmugam,
Exact Sciences Corp. Employment, Stock.
K. Cortes,
Exact Sciences Corp. Employment, Stock.
E. Chang,
Exact Sciences Corp. Employment, Stock.
P. J. Uren,
Exact Sciences Corp. Employment, Stock.
Biora Biosciences Stock.
A. Mazloom,
Exact Sciences Corp. Employment, Stock.
J. Garces,
Exact Sciences Corp. Employment, Stock.
G. A. Silvestri,
Nucleix Inc. ).
Delfi Diagnostics ).
Biodesix ), Other, Consulting.
Freenome ), Other, Consulting.
Candel therapeutics Other, Advisory Board.
D. W. Chesla, None.
R. W. Given,
Bayer Other, Speakers Bureau Service.
Johnson & Johnson Other, Speakers Bureau Service.
Francis Medical ), Travel.
MDX Health ).
Levee Medical ).
Dendreon ).
T. M. Beer,
Exact Sciences Corp. Employment, Stock.
Osteologic Stock.
Osheru Stock.
AstraZeneca Other, Consulting/Advisory Services.
F. Diehl,
Exact Sciences Corp. Employment, Stock.