scRNA——细胞类型自动注释

基于SCSA注释

SCSA是一种统计学模型注释，使用的数据是s4d8_quality_control.h5ad。（我也是抄别人的，尽量和教程保持一致，如果没一致，我就自己搞了）

数据—预处理—聚类

老规矩，该进行聚类还是得进行聚类。

adata = sc.read(
    filename="s4d8_quality_control.h5ad",
    backup_url="https://figshare.com/ndownloader/files/40014331",
)
# 因为它已经质控过了，所以直接预处理好了
ov.pp.preprocess(adata,mode='shiftlog|pearson',n_HVGs=2000,)
adata.raw = adata
adata = adata[:, adata.var.highly_variable_features]
ov.pp.scale(adata)
ov.pp.pca(adata, layer='scaled', n_pcs=50)
sc.pp.neighbors(adata, n_neighbors=15, n_pcs=30, use_rep='scaled|original|X_pca')
sc.tl.leiden(adata, key_added="leiden_res1", resolution=2.0)
adata.obsm["X_mde"] = ov.utils.mde(adata.obsm["scaled|original|X_pca"])

基于SCSA数据库注释

由于作者使用的pandas版本是1.6几的版本，太老了，在github 上有进行对应的更新（只要更新omicverse到1.6.10版本就行了），同样的也更新了pySCSA_2024_v1_plus.db这个文件，可以从谷歌云下载

scsa = ov.single.pySCSA(adata=adata,
                        foldchange=1.5,
                        pvalue=0.01,
                        celltype='normal',  # 正常的单细胞转录的话,使用normal, 癌细胞的话使用cancer
                        target='cellmarker',
                        # 正常的单细胞转录的话,使用cellmarker或是panglaodb, 癌细胞的话使用cancersea, 小鼠无法使用panglaodb
                        tissue='All',
                        species='Human',
                        model_path=r'pySCSA_2024_v1_plus.db'
)
scsa.cell_anno(clustertype='leiden_res1',
                    cluster='all', rank_rep=True)
scsa.cell_auto_anno(adata, key='scsa_celltype_cellmarker', clustertype='leiden_res1')
ov.utils.embedding(adata,
                   basis='X_mde',
                   color=['leiden_res1', 'scsa_celltype_cellmarker'],
                   legend_loc='on data',
                   frameon='small',
                   legend_fontoutline=1,
                   palette=ov.utils.palette()[14:],
                   show=True,  # show为False时,返回图片对象
)

如果需要获取指定细胞类型的marker基因，可以通过下面的代码进行获取。

marker_dict=ov.single.get_celltype_marker(adata,clustertype='scsa_celltype_cellmarker')
print(marker_dict)

# {'B cell': ['AFF3', 'BACH2', 'CD74', 'EBF1', 'BANK1', 'RALGPS2', 'IGHM', 'FCHSD2', 'LYN', 'ZCCHC7'], 'Hematopoietic stem cell': [], 'Monocyte': ['NAMPT'], 'Naive CD8+ T cell': [], 'Natural killer T (NKT) cell': [], 'Natural killer cell': ['B2M', 'GNLY'], 'Red blood cell (erythrocyte)': [], 'T cell': [], 'Unknown': []}

基于CellTypist的细胞类型注释

深度学习模型的自动注释方法CellTypist，可以直接注释出细胞的亚群，对于不认识的细胞类型，其会选择性注释，不会瞎注释，这与SCSA是相同的。

import omicverse as ov
import scanpy as sc
import celltypist
from celltypist import models
# 使用的数据还是s4d8_quality_control.h5ad
adata = sc.read(
        filename="data/s4d8_quality_control.h5ad",)

由于Celltypist要求使用的归一化值为10,000，而omicverse自带预处理的归一化值为500,000。所以我们提取原始counts进行还原，然后重新归一化。

adata=adata.raw.to_adata()
ov.utils.retrieve_layers(adata,layers='counts')
adata=ov.pp.preprocess(adata,mode='shiftlog|pearson',n_HVGs=2000,target_sum=1e4)

CellTypist模型预加载

机器学习的通病，需要下载模型到本地！烦死了！

models.download_models(force_update = True)

模型很小，才76MB，随便放！现在（2025.1.13）总共有54个模型。

💾 Downloading model [1/54]: Immune_All_Low.pkl 
💾 Downloading model [2/54]: Immune_All_High.pkl 
💾 Downloading model [3/54]: Adult_COVID19_PBMC.pkl
💾 Downloading model [4/54]: Adult_CynomolgusMacaque_Hippocampus.pkl
💾 Downloading model [5/54]: Adult_Human_MTG.pkl
💾 Downloading model [6/54]: Adult_Human_PancreaticIslet.pkl
💾 Downloading model [7/54]: Adult_Human_PrefrontalCortex.pkl
💾 Downloading model [8/54]: Adult_Human_Skin.pkl
💾 Downloading model [9/54]: Adult_Human_Vascular.pkl
💾 Downloading model [10/54]: Adult_Mouse_Gut.pkl
💾 Downloading model [11/54]: Adult_Mouse_OlfactoryBulb.pkl
💾 Downloading model [12/54]: Adult_Pig_Hippocampus.pkl
💾 Downloading model [13/54]: Adult_RhesusMacaque_Hippocampus.pkl
💾 Downloading model [14/54]: Autopsy_COVID19_Lung.pkl
💾 Downloading model [15/54]: COVID19_HumanChallenge_Blood.pkl
💾 Downloading model [16/54]: COVID19_Immune_Landscape.pkl
💾 Downloading model [17/54]: Cells_Adult_Breast.pkl
💾 Downloading model [18/54]: Cells_Fetal_Lung.pkl
💾 Downloading model [19/54]: Cells_Human_Tonsil.pkl
💾 Downloading model [20/54]: Cells_Intestinal_Tract.pkl
💾 Downloading model [21/54]: Cells_Lung_Airway.pkl
💾 Downloading model [22/54]: Developing_Human_Brain.pkl
💾 Downloading model [23/54]: Developing_Human_Gonads.pkl
💾 Downloading model [24/54]: Developing_Human_Hippocampus.pkl
💾 Downloading model [25/54]: Developing_Human_Organs.pkl
💾 Downloading model [26/54]: Developing_Human_Thymus.pkl
💾 Downloading model [27/54]: Developing_Mouse_Brain.pkl
💾 Downloading model [28/54]: Developing_Mouse_Hippocampus.pkl
💾 Downloading model [29/54]: Fetal_Human_AdrenalGlands.pkl
💾 Downloading model [30/54]: Fetal_Human_Pancreas.pkl
💾 Downloading model [31/54]: Fetal_Human_Pituitary.pkl
💾 Downloading model [32/54]: Fetal_Human_Retina.pkl
💾 Downloading model [33/54]: Fetal_Human_Skin.pkl
💾 Downloading model [34/54]: Healthy_Adult_Heart.pkl
💾 Downloading model [35/54]: Healthy_COVID19_PBMC.pkl
💾 Downloading model [36/54]: Healthy_Human_Liver.pkl
💾 Downloading model [37/54]: Healthy_Mouse_Liver.pkl
💾 Downloading model [38/54]: Human_AdultAged_Hippocampus.pkl
💾 Downloading model [39/54]: Human_Colorectal_Cancer.pkl
💾 Downloading model [40/54]: Human_Developmental_Retina.pkl
💾 Downloading model [41/54]: Human_Embryonic_YolkSac.pkl
💾 Downloading model [42/54]: Human_Endometrium_Atlas.pkl
💾 Downloading model [43/54]: Human_IPF_Lung.pkl
💾 Downloading model [44/54]: Human_Longitudinal_Hippocampus.pkl
💾 Downloading model [45/54]: Human_Lung_Atlas.pkl
💾 Downloading model [46/54]: Human_PF_Lung.pkl
💾 Downloading model [47/54]: Human_Placenta_Decidua.pkl
💾 Downloading model [48/54]: Lethal_COVID19_Lung.pkl
💾 Downloading model [49/54]: Mouse_Dentate_Gyrus.pkl
💾 Downloading model [50/54]: Mouse_Isocortex_Hippocampus.pkl
💾 Downloading model [51/54]: Mouse_Postnatal_DentateGyrus.pkl
💾 Downloading model [52/54]: Mouse_Whole_Brain.pkl
💾 Downloading model [53/54]: Nuclei_Lung_Airway.pkl
💾 Downloading model [54/54]: Pan_Fetal_Human.pkl

感兴趣的可以从models.json中了解更详细的内容。（太多了，根本整不过来！）

模型预测

predictions = celltypist.annotate(adata, model='Immune_All_Low.pkl', majority_voting=True, mode='prob match',
                                      p_thres=0.5)
adata_new = predictions.to_adata()
majority_fig = ov.utils.embedding(adata_new,
                                      basis='X_mde',
                                      color=["majority_voting"],
                                      # title=['Cell type'],
                                      palette=sc.pl.palettes.default_102,
                                      show=False,
                                      frameon='small',
                                      wspace=0.45,
                                      legend_fontsize = 8,
                                      legend_fontoutline=1)

大部分细胞类型被标注为Unassigned，这是由于我们设定了预测概率0.5作为阈值，低于该阈值的细胞不被注释，表明Celltypist在注释复杂的未知的数据上的效果，并不如手动注释或者是SCSA！

使用人工智能注释

人工智能仅能注释细胞大类，不能注释细胞亚型，还是有缺陷的。先是使用get_celltype_marker获取这这写簇的特征基因是什么。

all_markers = ov.single.get_celltype_marker(adata, clustertype='leiden', rank=True,
                                                key='rank_genes_groups',
                                                foldchange=2, topgenenumber=5)
# 假设有这些
{'0': ['LDHB', 'CD3D', 'NOSIP', 'CD3E', 'CD7'],
 '1': ['LTB', 'LDHB', 'IL32', 'IL7R', 'CD3D'], 
 '10': ['PPBP', 'PF4', 'GNG11', 'SPARC', 'SDPR'], 
 '2': ['LYZ', 'S100A9', 'S100A8', 'FCN1', 'TYROBP'],
 '3': ['CD74', 'CD79A', 'HLA-DRA', 'CD79B', 'HLA-DPB1'], 
 '4': ['CCL5', 'NKG7', 'CST7', 'GZMA', 'IL32'], 
 '5': ['LST1', 'FCER1G', 'COTL1', 'AIF1', 'IFITM3'], 
 '6': ['NKG7', 'GZMB', 'GNLY', 'PRF1', 'CTSW'], 
 '7': ['CCL5', 'CD3D', 'IL32', 'GZMK'], 
 '8': ['FTL', 'S100A8', 'FTH1', 'TYROBP', 'S100A9'], 
 '9': ['HLA-DPA1', 'HLA-DRB1', 'HLA-DPB1', 'HLA-DRB5', 'HLA-DRA']}

然后获取api-key，omicverse支持openai， 通义千问，Kimi等人工智能大厂的api调用。

import os
os.environ['AGI_API_KEY'] = 'sk-****'
# api-key需要自己生成，半个小时不到就可以搞定了
# 然后将上面的all_markers发给人工智能
result = ov.single.gptcelltype(all_markers, tissuename='PBMC', speciename='human',
                                   model='qwen-plus', provider='qwen',
                                   topgenenumber=5)

假设返回的是下面的内容

{	'0': 'T cells  ', 
	'1': 'T cells  ', 
	'10': 'Megakaryocytes/Platelets  ', 
	'2': 'Monocytes/Macrophages  ', 
	'3': 'B cells  ', 
	'4': 'Natural Killer Cells  ', 
	'5': 'Dendritic Cells/Monocytes  ',
	'6': 'Natural Killer Cells  ', 
	'7': 'T cells/Natural Killer Cells  ', 
	'8': 'Monocytes/Macrophages  ', 
	'9': 'B cells'}
	

只要把这些内容注释到里面数据里面就可以了

adata.obs['gpt_celltype'] = adata.obs['leiden'].map(result).astype('category')
adata.write_h5ad(output_path)