Hugging Face's logo

Spaces:
KeTuTu
/
OV_Agentic_EXP_SambaNova
Sleeping

App Files Community
OV_Agentic_EXP_SambaNova
File size: 6,535 Bytes 
2999286
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
 
#!/usr/bin/env python
# coding: utf-8

# # Spatial integration and clustering
# 
# STAligner is designed for alignment and integration of spatially resolved transcriptomics data.
# 
# STAligner first normalizes the expression profiles for all spots and constructs a spatial neighbor network using the spatial coordinates. STAligner further employs a graph attention auto-encoder neural network to extract spatially aware embedding, and constructs the spot triplets based on current embeddings to guide the alignment process by attracting similar spots and discriminating dissimilar spots across slices. STAligner introduces the triplet loss to update the spot embedding to reduce the distance from the anchor to positive spot, and increase the distance from the anchor to negative spot. The triplet construction and auto-encoder training are optimized iteratively until batch-corrected embeddings are generated. STAligner can be applied to integrate ST datasets to achieve alignment and simultaneous identification of spatial domains from different biological samples in (a), technological platforms (I), developmental (embryonic) stages (II), disease conditions (III) and consecutive slices of a tissue for 3D slice alignment (IV
# 
# Zhou, X., Dong, K. & Zhang, S. Integrating spatial transcriptomics data across different conditions, technologies and developmental stages. Nat Comput Sci 3, 894–906 (2023). https://doi.org/10.1038/s43588-023-00528-w
# 
# ![image.png](attachment:00790548-59f9-4fad-a1e3-a52c3ae98d44.png)).

# In[1]:


from scipy.sparse import csr_matrix
import omicverse as ov
import scanpy as sc
import anndata as ad
import pandas as pd
import os

ov.utils.ov_plot_set()


# # Preprocess data
# 
# Here, We use the mouse olfactory bulb data generated by Stereo-seq and Slide-seqV2. The processed Stereo-seq and Slide-seqV2 data can be downloaded from https://drive.google.com/drive/folders/1Omte1adVFzyRDw7VloOAQYwtv_NjdWcG?usp=share_link. and the original tutorals can be finded from https://staligner.readthedocs.io/en/latest
# 
# Here is a critical point that must be clarified: for STAligner, it first calculates highly variable genes before concating annadata samples. Therefore, the number of highly variable genes should not be selected too low. Otherwise, in the case of a large number of samples, the downstream features for STAligner training would be insufficient, impacting the model's performance.
# 
# When using STAligner, it is necessary to adjust the **rad_cutoff** parameter according to different data to ensure that each spot has an **average of 5-10 adjacent spots** connected to it. Such as: "11.3356 neighbors per cell on average."
# 

# In[2]:


Batch_list = []
adj_list = []
section_ids = ['Slide-seqV2_MoB', 'Stereo-seq_MoB']
print(section_ids)
pathway = '/storage/zengjianyangLab/hulei/scRNA-seq/scripts/STAligner'

for section_id in section_ids:
    print(section_id)
    adata = sc.read_h5ad(os.path.join(pathway,section_id+".h5ad"))

    # check whether the adata.X is sparse matrix
    if isinstance(adata.X, pd.DataFrame):
        adata.X = csr_matrix(adata.X)
    else:
        pass

    adata.var_names_make_unique(join="++")

    # make spot name unique
    adata.obs_names = [x+'_'+section_id for x in adata.obs_names]

    # Constructing the spatial network
    ov.space.Cal_Spatial_Net(adata, rad_cutoff=50) # the spatial network are saved in adata.uns[‘adj’]

    # Normalization
    sc.pp.highly_variable_genes(adata, flavor="seurat_v3", n_top_genes=10000)
    sc.pp.normalize_total(adata, target_sum=1e4)
    sc.pp.log1p(adata)

    adata = adata[:, adata.var['highly_variable']]
    adj_list.append(adata.uns['adj'])
    Batch_list.append(adata)


# In[3]:


Batch_list


# In[4]:


adata_concat = ad.concat(Batch_list, label="slice_name", keys=section_ids)
adata_concat.obs["batch_name"] = adata_concat.obs["slice_name"].astype('category')
print('adata_concat.shape: ', adata_concat.shape)


# # Training STAligner model
# 
# Here, we used `ov.space.pySTAligner` to construct a STAGATE object to train the model.
# 
# We are using the `train_STAligner_subgraph` function from STAligner to reduce GPU memory usage, each slice is considered as a subgraph for training.

# In[5]:


get_ipython().run_cell_magic('time', '', "# iter_comb is used to specify the order of integration. For example, (0, 1) means slice 0 will be algined with slice 1 as reference.\niter_comb = [(i, i + 1) for i in range(len(section_ids) - 1)]\n\n# Here, to reduce GPU memory usage, each slice is considered as a subgraph for training.\nSTAligner_obj = ov.space.pySTAligner(adata_concat, verbose=True, knn_neigh = 100, n_epochs = 600, iter_comb = iter_comb,\n                                     batch_key = 'batch_name',  key_added='STAligner', Batch_list = Batch_list)\n")


# In[6]:


STAligner_obj.train()


# We stored the latent embedding in `adata.obsm['STAligner']`.

# In[7]:


adata = STAligner_obj.predicted()


# # Clustering the space
# 
# We can use `GMM`, `leiden` or `louvain` to cluster the space.
# 
# `ov.utils.cluster(adata,use_rep='STAligner',method='GMM',n_components=7,covariance_type='full', tol=1e-9, max_iter=1000, random_state=3607`
# 
# or `sc.pp.neighbors(adata, use_rep='STAligner', random_state=666)`            
# `ov.utils.cluster(adata,use_rSTAlignerGATE',method='leiden',resolution=1)`

# In[8]:


sc.pp.neighbors(adata, use_rep='STAligner', random_state=666)
ov.utils.cluster(adata,use_rep='STAligner',method='leiden',resolution=0.4)
sc.tl.umap(adata, random_state=666)
sc.pl.umap(adata, color=['batch_name',"leiden"],wspace=0.5)


# We can also map the clustering results back to the original spatial coordinates to obtain spatially specific clustering results.

# In[9]:


import matplotlib.pyplot as plt
spot_size = 50
title_size = 15
fig, ax = plt.subplots(1, 2, figsize=(6, 3), gridspec_kw={'wspace': 0.05, 'hspace': 0.2})
_sc_0 = sc.pl.spatial(adata[adata.obs['batch_name'] == 'Slide-seqV2_MoB'], img_key=None, color=['leiden'], title=['Slide-seqV2'],
                      legend_fontsize=10, show=False, ax=ax[0], frameon=False, spot_size=spot_size, legend_loc=None)
_sc_0[0].set_title('Slide-seqV2', size=title_size)

_sc_1 = sc.pl.spatial(adata[adata.obs['batch_name'] == 'Stereo-seq_MoB'], img_key=None, color=['leiden'], title=['Stereo-seq'],
                      legend_fontsize=10, show=False, ax=ax[1], frameon=False, spot_size=spot_size)
_sc_1[0].set_title('Stereo-seq',size=title_size)
_sc_1[0].invert_yaxis()
plt.show()


# In[ ]: