10X单细胞scanpy bbknn算法做多样本，不同批次整合

参考：http://events.jianshu.io/p/d855ee1a121f

scanpy做10X单细胞多样本整合与Seurat差别比较大，很多学者喜欢使用seurat来进行整合，但实际情况是基于R平台的Seurat MNN 整合算法非常消耗内存，并不适合进行多样本大量批次的整合

首先来看看简介，The ingest function assumes an annotated reference dataset that captures the biological variability of interest.（这一句话就很重要，首先需要一个注释好的参考数据集，然后来“捕获”疾病样本的生物学变化） 。The rational（理论） is to fit amodel on the reference data and use it to project new data（用参考数据集拟合一个model，从而来插入新的数据集）. For the time being, this model is a PCA combined with a neighbor lookup search tree, for which we use UMAP’s implementation

这块儿需要注意了，做整合分析的时候，scanpy需要我们有一个注释好的data作为reference，从理性上讲应该这样，但很多时候我们用Seurat做整合用不到这个，个人欣赏scanpy的做法，更为理性。

scanpy给出的官方示例

We refer to this asymmetric dataset integration as ingesting annotations from an annotated reference adata_ref into an adata that still lacks this annotation（跟上面介绍的一致）. It is different from learning a joint representation that integrates

datasets in a symmetric way as BBKNN, Scanorma, Conos, CCA（CCA是Seurat做整合常用的方法） (e.g. in Seurat) or a conditional VAE (e.g. in scVI, trVAE) would do,but comparable to the initiall MNN implementation in scran(scran这个软件目前用的也比较少,更多的是需要这个软件其中一部分的功能) 不过scanpy做样本整合的思路和方法确实有别于其他的软件。

加载模块和示例数据

import scanpy as sc 
import pandas as pd 
import seaborn as sns 
adata_ref = sc.datasets.pbmc3k_processed()  # this is an earlier version of the dataset from t
adata = sc.datasets.pbmc68k_reduced() 

第一个值得注意的地方

var_names=adata_ref.var_names.intersection(adata.var_names)   #adata.var_names指高变基因的名字
adata_ref=adata_ref[:, var_names] #预先经过排序了吗就开始切片？
adata=adata[:,var_names] 

这个地方需要一点小心，首先高变基因是adata的，但是对于ref数据集并不再计算高变基因，而是直接用adata的高变基因与ref的所有基因取交集，然后对两个数据集进行切割，保留分析用到的数据，这个地方需要注意，因为涉及到很多的问题，比如高变基因的数量，挑选阈值等等，个人也欣赏这个方法，可以在降维聚类的过程中体现出正常和疾病数据集的差异。对ref进行降维聚类,疾病样本的高变基因提取ref数据集，然后对数据集进行降维聚类，

sc.pp.pca(adata_ref) 
sc.pp.neighbors(adata_ref) 
sc.tl.umap(adata_ref) 
sc.pl.umap(adata_ref, color='louvain')

image.png

图片不是ref本身的高变基因进行降维聚类的结果,因为不同细胞类型之间有混合,这样的目的就是在与整合疾病数据集

转换标签

Let’s map labels and embeddings from adata_ref to adata based on a chosen representation. Here, we use adata_ref.obsm['X_pca'] to map cluster labels and the UMAP coordinates.

sc.tl.ingest(adata, adata_ref, obs='louvain') 
adata.uns['louvain_colors'] = adata_ref.uns['louvain_colors'] # fix colors 
sc.pl.umap(adata, color=['louvain', 'bulk_labels'], wspace=0.5)

image.png

注意看这里你会发现adata的细胞类型和cluster和adata_ref很像，这个在文章的后面会再进行解释

进行数据整合

adata_concat = adata_ref.concatenate(adata, batch_categories=['ref', 'new']) 
adata_concat.obs.louvain = adata_concat.obs.louvain.astype('category') 
adata_concat.obs.louvain.cat.reorder_categories(adata_ref.obs.louvain.cat.categories, inplace=True) # fix category ordering 
adata_concat.uns['louvain_colors'] = adata_ref.uns['louvain_colors'] # fix category colors
sc.pl.umap(adata_concat, color=['batch', 'louvain'])

image.png

While there seems to be some batch-effect in the monocytes and dendritic cell clusters, the new data is otherwise mapped relatively homogeneously.存在一定的批次效应

The megakaryoctes are only present in adata ref and no cells from adata map onto them. If interchanging reference data and query data, Megakaryocytes do not appear as a separate cluster anymore. This is an extreme case as the reference data is very small; but one should always question if the reference data contain enough biological variation to meaningfully accomodate query data.

整合数据进行批次矫正

sc.tl.pca(adata_concat) 
sc.external.pp.bbknn(adata_concat, batch_key='batch') 
sc.tl.umap(adata_concat) 
sc.pl.umap(adata_concat, color=['batch', 'louvain']) 

image.png

bbknn的问题在于，如果adata数据集中有新的细胞类型，通过这种映射方法是很难找到的。

Density plot（核密度图）

sc.tl.embedding_density(adata_concat, groupby='batch') 
sc.pl.embedding_density(adata_concat, groupby='batch')

image.png

Partial visualizaton of a subset of groups in embedding

for batch in ['batch1', 'batch2', 'batch3']: 
    sc.pl.umap(adata_concat, color='batch', groups=[batch])

image.png

image.png

image.png