Integrate RNA+ATAC in mouse brain section

In this notebook, we used SpaMosaic to integrate RNA and ATAC modalities from one postnatal mouse brain section. This section was measured by spatial-RNA-epigenome seq (Zhang et al., Nature. 2023). It consists of the transcriptome and ATAC assays of a P21 mouse brain section with a capture array of 50×50 pixels, each 20 µm in diameter.

Data used in this notebook can be downloaded from google drive.

import os
import scanpy as sc
from os.path import join

from spamosaic.framework import SpaMosaic

os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'  # for CuBLAS operation and you have CUDA >= 10.2
import spamosaic.utils as utls
from spamosaic.preprocessing import RNA_preprocess, ADT_preprocess, Epigenome_preprocess

ad1_rna  = sc.read_h5ad('./s1_adata_rna.h5ad')
ad1_atac = sc.read_h5ad('./s1_adata_atac.h5ad')

preprocessing

SpaMosaic expects batch-corrected low-dimensional representations for each modality as input. Therefore, for the raw count assays of each modality, SpaMosaic performs feature selection and dimension reduction to obtain the low-dimensional representations. Then, Harmony is performed to integrate these representations in each modality (optional, depending on the presence of strong batch effects).

By default, the feature selection and dimension reduction consist of following steps:

• RNA assays: highly variable gene selection \(\rightarrow\) log normalization \(\rightarrow\) PCA

• epigenome assays: highly variable peak selection \(\rightarrow\) TF-IDF transformation \(\rightarrow\) log normalization \(\rightarrow\) PCA

• protein assays: centered log ratio normalization \(\rightarrow\) PCA

In vertical integration, SpaMosaic requires the input in the format of

{
    'modality1_name': [adata_mod1],
    'modality2_name': [adata_mod2],
    ...
}

In the dictionary, each key represents a modality and each modality key corresponds to list of anndata objects. Each anndata object contains modality-specific information for a particular section.

input_dict = {
    'rna':  [ad1_rna, ],
    'atac': [ad1_atac,]
}

input_key = 'dimred_bc'

Parameters in the following RNA_preprocess and Epigenome_preprocess function:

• batch_corr: whether to perform batch correction with Harmony

• n_hvg/n_peak: how many highly variable genes/peaks to select

• batch_key: the key in the row metadata that holds the batch labels

• key: the key in .obsm that will hold the output low-dimensional representations

After preprocessing, the low-dimensional representations can be accessed by: ad1_rna.obsm[input_key], ad1_atac.obsm[input_key].

RNA_preprocess(input_dict['rna'], batch_corr=False, n_hvg=5000, batch_key='src', key=input_key)
Epigenome_preprocess(input_dict['atac'], batch_corr=False, n_peak=50000, batch_key='src', key=input_key)

training

SpaMosaic employs modality-specific graph neural networks to embed each modality’s input into latent space. In horizontal integration, all sections have only one modality, thus each section has only one set of embeddings.

The crucial parameters include:

• intra_knns: how many nearest neighbors to consider when searching for spatial neighbors within each section (list or integer)

• inter_knn_base: how many nearest neighbors to consider when searching for mutual nearest neighbors between sections (integer)

• w_g: the weight for spatial-adjacency graph

The following parameters are recommended to use in complex integration scenarios, like varying resolution or size across sections

• smooth_input: whethere to smooth the input representations (bool)

• smooth_L: number of LGCN layers for smoothing input

• inter_auto_knn: whether to automatically balance the kNN during MNN searching (bool)

• rmv_outlier: whether to remove outlier of MNN (bool)

• contamination: percentage of removed MNN outlier

for training:

• net: which graph neural network to use (only support wlgcn now)

• lr: learning rate

• n_epochs: number of training epochs

• w_rec_g: the loss weight for reconstructing original graph structure. If the target dataset contains protein modality, we recommend a low value for w_rec_g (e.g., 0); if it contains epigenome modality, we recommend a high value for w_rec_g (e.g., 1).

model = SpaMosaic(
    modBatch_dict=input_dict, input_key=input_key,
    batch_key='src', intra_knns=10,
    seed=1234,
    device='cuda:0'
)

model.train(net='wlgcn', lr=0.01, T=0.01, n_epochs=100)

batch0: ['rna', 'atac']
------Calculating spatial graph...
The graph contains 23720 edges, 2372 cells.
10.0000 neighbors per cell on average.
------Calculating spatial graph...
The graph contains 23720 edges, 2372 cells.
10.0000 neighbors per cell on average.
Searching MNN within rna
Number of mnn pairs for rna:0
Searching MNN within atac
Number of mnn pairs for atac:0

100%|███████████████████████████████████████████████████████████████████████████████████| 100/100 [00:02<00:00, 45.38it/s]

inference

After training, SpaMosaic can infer the modality-specific embedding for each section. These embeddings will be directly saved in original anndata objects. For example, the RNA-specific embeddings can be accessed by ad1_rna.obsm['emb'], the ATAC-specific embeddings can be accessed by ad1_atac.obsm['emb'].

The following infer_emb function will return a new list of anndata objects, which save the final/merged embeddings for spatial clustering. The final embeddings can be accessed by ad_embs[0].obsm['merged_emb'].

In vertical integration, there is only one section, and the final merged embeddings can be obtained by averaging or concatenating the modality-specific embeddings.

ad_embs = model.infer_emb(input_dict, emb_key='emb', final_latent_key='merged_emb')
ad_mosaic = sc.concat(ad_embs)
ad_mosaic = utls.get_umap(ad_mosaic, use_reps=['merged_emb'])

/home/yanxh/anaconda3/envs/spamosaic-env/lib/python3.8/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from .autonotebook import tqdm as notebook_tqdm

Visualize the final embeddings using UMAP plots, spot are colored by batch labels

utls.plot_basis(ad_mosaic, basis='merged_emb_umap', color=['src'])

clustering

Parameters for the following utls.clustering function:

• n_cluster: expected number of clusters

• used_obsm: the target key in the ad_mosaic.obsm that holds the embeddings for clustering input

• algo: which clustering algorithm to use, mclust or kmeans. Sometimes mclust can fail to output clustering, it will automatically output the kmeans outputs.

• key: the key in the row metadata that will hold the output clustering labels.

Generally, SpaMosaic works better with the mclust algorithm in spatial domain identification task. The final/merged embeddings are used as input for mclust and the clustering labels can be accessed by ad_mosaic.obs['mclust'].

utls.clustering(ad_mosaic, n_cluster=6, used_obsm='merged_emb', algo='mclust', key='mclust')
utls.split_adata_ob(ad_embs, ad_mosaic, 'obs', 'mclust')

R[write to console]:                    __           __
   ____ ___  _____/ /_  _______/ /_
  / __ `__ \/ ___/ / / / / ___/ __/
 / / / / / / /__/ / /_/ (__  ) /_
/_/ /_/ /_/\___/_/\__,_/____/\__/   version 6.0.0
Type 'citation("mclust")' for citing this R package in publications.

fitting ...
  |======================================================================| 100%

Spatial plots of clustering label for individual sections

for ad in ad_embs:
    utls.plot_basis(ad, 'spatial', 'mclust', s=100)