INSTINCT

Multi-sample integration of spatial chromatin accessibility sequencing data via stochastic domain translation

System requirements

The package development version is tested on Windows operating systems. The developmental version of the package has been tested on the following systems:

Linux: Ubuntu 20.04
Windows

Installation

Clone the repository.

git clone https://github.com/yyLIU12138/INSTINCT.git
cd INSTINCT

Create an environment.

conda create -n epi_INSTINCT python=3.10
conda activate epi_INSTINCT

Install the required packages.

pip install -r requirement.txt

Install INSTINCT.

python setup.py build
python setup.py install

Installation takes a few minutes.

Tutorial

Detailed version of tutorials for INSTINCT can be found on the Read the Docs website.

Import the package.

import torch
import anndata as ad
from sklearn.decomposition import PCA
import INSTINCT
import warnings
warnings.filterwarnings("ignore")

Load the anndata type data samples into a list.

data_dir = '../demo_data/EpiTran_MouseBrain_Jiang2023/'
slice_name_list = ['E11_0-S1', 'E13_5-S1', 'E15_5-S1', 'E18_5-S1']
cas_list = [ad.read_h5ad(data_dir + sample + '_atac.h5ad') for sample in slice_name_list]
for j in range(len(cas_list)):
    cas_list[j].obs_names = [x + '_' + slice_name_list[j] for x in cas_list[j].obs_names]

Merge the peaks.

cas_list = INSTINCT.peak_sets_alignment(cas_list)

Preprocessing (If the data samples already incorparate fragment count matrices, then set use_fragment_count=False).

adata_concat = ad.concat(cas_list, label="slice_name", keys=slice_name_list)
INSTINCT.preprocess_CAS(cas_list, adata_concat, use_fragment_count=True, min_cells_rate=0.03)

Use PCA to reduce the dimensionality of the concatenated data to 100. The matrix of shape N*100 should be stored in adata_concat.obsm['X_pca'].

pca = PCA(n_components=100, random_state=1234)
input_matrix = pca.fit_transform(adata_concat.X.toarray())
adata_concat.obsm['X_pca'] = input_matrix

Construct the neighbor graph

INSTINCT.create_neighbor_graph(cas_list, adata_concat)

Train the model. The low-dimensional representations for spots are stored in .obsm['INSTINCT_latent'] of each slice in cas_list.

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
INSTINCT_model = INSTINCT.INSTINCT_Model(cas_list,
                                         adata_concat,
                                         input_mat_key='X_pca',  # the key of the input matrix in adata_concat.obsm
                                         input_dim=100,  # the input dimension
                                         hidden_dims_G=[50],  # hidden dimensions of the encoder and the decoder
                                         latent_dim=30,  # the dimension of latent space
                                         hidden_dims_D=[50],  # hidden dimensions of the discriminator
                                         lambda_adv=1,  # hyperparameter for the adversarial loss
                                         lambda_cls=10,  # hyperparameter for the classification loss
                                         lambda_la=20,  # hyperparameter for the latent loss
                                         lambda_rec=10,  # hyperparameter for the reconstruction loss
                                         seed=1236,  # random seed
                                         learn_rates=[1e-3, 5e-4],  # learning rate
                                         training_steps=[500, 500],  # training_steps
                                         use_cos=True,  # use cosine similarity to find the nearest neighbors
                                         margin=10,  # the margin of latent loss
                                         alpha=1,  # the hyperparameter for triplet loss
                                         k=50,  # the amount of neighbors to find
                                         device=device)

INSTINCT_model.train(report_loss=True, report_interval=100)

INSTINCT_model.eval(cas_list)

Training the model takes about one minute using GPU (RTX 4090D 24GB).