github主页

TCRGP is a novel Gaussian process method that can predict if TCRs recognize certain epitopes. This method can utilize different CDR sequences from both TCRα and TCRβ chains from single-cell data and learn which CDRs are important in recognizing the different epitopes.

• It is well known that the CDR3β of a TCR is important in recognizing peptides presented to the T cell.
• We propose a method called TCRGP which builds on non-parametric modelling using Gaussian process (GP) classification. The probabilistic formulation of GPs allows robust model inference already from small data sets, which is a great benefit as currently there exists very limited amounts of reported TCR-epitope interactions in curated databases.

一、下载

安装好tensorflow和GPflow
如果网络太慢，可以使用镜像安装
直接下载github上面整个的zip，并解压使用

ATTENTION!!!!!!!!!!!
必须按照

to use TCRGP, you will need to have

• TensorFlow (We have used version 1.8.0)---对应Python为3.6（最好在对应的虚拟环境安装）
• GPflow (We have used version 1.1.1)

tensorflow1.8.0安装可以参考这个
)

关于如何在linux jupyter notebook使用conda虚拟环境,使用的方法二
下载好后在对应Conda环境中输入jupyter notebook，然后输入对应网址即可

如果报错ImportError: cannot import name 'secure_write'

二、导入

将jupyter notebook的默认读取路径设置到上述tcrgp解压位置

%pwd #获得路径
%cd #更改路径
import tcrgp

注意，对于虚拟环境中安装包的位置应指定到对应路径，如pip3 install matplotlib -i https://pypi.tuna.tsinghua.edu.cn/simple/ -t /home/user/test/miniconda3/envs/tfpy3/lib/python3.6/site-packages

查看安装包的位置：

import tensorflow
print(tensorflow.__path__)

三、训练集获得

VDJdb

如图，导出tsv文件

进一步筛选：

1. confidence score of at least 1
2. 选择所需的物种，如小鼠
3. 选择所需的HLA亚型。对于人类，可以看HLA-A*02，对于小鼠，可以筛选所需品系的小鼠。

比如，对于普通的C57小鼠，其MHC-I为H-2Kb和H-2Db，其MHC-II分子为I-A，因此，我需要筛选H-2Kb和H-2Db提呈的抗原肽

4. 至少存在一条识别序列包括50条TCRB序列

导出为tsv文件

四、模型构建及验证

按照参考走
模型构建后验证方法：

Leave-one-subject-out cross-validations
Leave-one-subject-out (loso) cross-validations can be used to evaluate the performance of the model.

解释

注意，在预测时需要注意

1. 修改lmax3，即最大CDR3的氨基酸数
2. TRBV12-2+TRBV13-2 中的+要改成;
3. 阈值为>0.85或0.9
   具体我再看看，有需要补充的我再更新