COVER
❝Sun Y, Selvarajan S, Zang Z, et al. Artificial intelligence defines protein-based classification of thyroid nodules [published correction appears in Cell Discov. 2022 Sep 30;8(1):100]. Cell Discov. 2022;8(1):85. Published 2022 Sep 6. doi:10.1038/s41421-022-00442-x
❞❝论文通讯为西湖大学郭天南和李子青，分别是组学和AI领域的大佬。
❞
Abstract
作者开发一种「AI-defined protein-based biomarker panel」用于甲状腺结节诊断分类。最初基于福尔马林固定石蜡包埋(formalin-fixed paraffin-embedded，FFPE)样本，进一步改进为微量细针穿刺(fine-needle aspiration，FNA)组织标本。作者首先开发了一个基于1724个「FFPE」甲状腺组织样本的蛋白质组的19个蛋白质生物标记的「神经网络模型」。该分类器在分类恶性甲状腺结节的发现集中达到91%以上的准确性。在288个结节的回顾性队列中，通过盲法分析对分类器进行了外部验证(准确率89%;FFPE)和来自12个独立临床中心的294例FNA活检(准确率85%)的前瞻性队列。
❝神经网络模型的代码作者已开源，有兴趣的同学可以自行去原文查看链接。当然，如果大家需要木舟对代码进行注释复现，可以多多「点赞」、「在看」和「分享」本文。
❞
The project design and workflow of the FFPE-PCT-DIA pipeline
image-20221113214941483
❝主要流程还是比较简单的，这里就不赘述。
❞
Feature selection and classifier development
其他部分都比较常规，直接看一下模型部分。
Schematic workflow
特征选择
Data splitting
如图，将「发现队列」（discovery dataset）分为2/3的datasetA和剩下的datasetB。在datasetA中使用三折交叉验证（three-fold cross-validation）进行特征筛选（Feature selection）。
Flow diagram of genetic algorithm for protein features selection
根据「差异分析结果」以及「已发表文献报道」或TCGA、OMIM数据库分析与甲状腺癌诊断有关的蛋白，将原始的6689个蛋白筛选到1302个。再根据样本缺失率<35%，得到767个蛋白质。
❝35%缺失率阈值怎么确定的？
In process of feature selection, we explored different screening conditions, using 25%,30%,35%, and 40% as thresholds for the missing value. The data showed that at more relaxed screening conditions the higher number of candidate features were nominated and that the results became more unstable in the independent validation set. The features with higher missing value rates, although potentially providing better classification were not stably detected and may negatively affect the quality of the model if such features are selected. Finally, in our classifier, we clearly defined the criteria for feature selection and used 767 proteins that were missing in less than 35% of samples
❞
随机生成500个每个包含19个蛋白（图中有误）的panel，在datasetA中进行交叉验证，并计算fitness。datasetB对性能进行评估，选出最优panel。
❝为了消除seed人为设置的随机性，作者比较了seed从0取到14对特征筛选的影响。
❞❝为什么选择19个蛋白：
To determine the count of protein features in the panel, we compare the model performances on the different number of features from four to 35 proteins. The more features the better accuracy and AUC achieved and reached a plateau when using 13 proteins . But for the most similar two histopathology types, FA and FTC, 19 proteins achieved the highest accuracy. Therefore, we used 19 proteins as a panel in the present study.
line chart showing the model performance on different counts of protein features❞
fitness计算公式：
❝作者还使用了Lasso和RF对特征进行筛选，发现性能没有本方法好。
❞
使用「SHapley Additive exPlanations」 (「SHAP」) 算法计算特征重要性。
he importance rank of the selected 19 protein features
神经网络模型训练
Structural diagram of neural network
使用全部发现队列进行训练。提取19个特征的表达矩阵。使用PyTorch进行MLP训练。使用5折交叉验证进行超参调优。
样本分类
对独立队列进行预测，评估模型性能。
代码
zangzelin/thyroid-project (github.com)(https://github.com/zangzelin/thyroid-project)
往期
CNS图表复现|生信分析|R绘图 资源分享&讨论群！
这图怎么画| 有点复杂的散点图
这图怎么画 | 相关分析棒棒糖图
组学生信| Front Immunol |基于血清蛋白质组早期诊断标志筛选的简单套路
(免费教程+代码领取)|跟着Cell学作图系列合集
Q&A | 如何在论文中画出漂亮的插图？
跟着 Cell 学作图 | 桑葚图(ggalluvial)
R实战 | Lasso回归模型建立及变量筛选
跟着 NC 学作图 | 互作网络图进阶(蛋白+富集通路)(Cytoscape)
R实战 | 给聚类加个圈圈(ggunchull)
R实战 | NGS数据时间序列分析(maSigPro)
跟着 Cell 学作图 | 韦恩图(ggVennDiagram)
