lixiangchun/miscell

Mining Information from Single-Cell high-throughput transcriptome data

Miscell

Mining Information from Single-Cell high-throughput transcriptome data. Some soure code was copied from facebookresearch/MoCo.

The following example was tested with python-3.9.2 and torch-1.7.1-cu110.

1. Installation

git clone https://github.com/lixiangchun/Miscell.git
cd Miscell

pip install torch
pip install sklearn
pip install scanpy

2. Training Miscell on own data

The model was trained on GPUs and at least two GPUs are required.

#!/bin/bash

# Model name: densenet11, densenet21, densenet29 and densenet63
arch=densenet11
# Number of input features
in_features=1686
# We recommend the user to normalize the feature before training the model
data=data/x_new_scanpy_HVG.npz

python main.py \
-a $arch -j 1 -b 256 --lr 0.24 \
--dist-url "tcp://localhost:10001" \
--world-size 1 \
--rank 0 \
--multiprocessing-distributed \
--outdir checkpoint \
--in_features $in_features \
--save_frequency 5 \
--mlp --moco-t 0.2 --moco-k 1024 --moco-m 0.999 --moco-dim 64 $data

3. Extract feature with pretrained model

import utils
import numpy as np

X = np.load("data/x_new_scanpy_HVG.npz")["x"]
in_features = 1686
assert in_features == X.shape[1]

checkpoint = "checkpoint_0199.pth.tar"

model = utils.load_pretrained_model(arch, in_features, checkpoint, return_feature=True)

features = utils.extract_features(model, X)

4. Perform DBSCAN clustering on t-SNE features obtained from pretrained model

Install FIt-SNE.

import sklearn.cluster
import sys
# Modify the path to FIt-SNE accordingly.
sys.path.append('/home/lixc/software/github/FIt-SNE/')
from fast_tsne import fast_tsne

tsne = fast_tsne(features, seed=123, nthreads=12, perplexity_list=[30, 36, 42, 48])
_, y = sklearn.cluster.dbscan(tsne, eps=1, min_samples=5, algorithm='auto')

5. Perform clustering with scanpy

# https://scanpy-tutorials.readthedocs.io/en/latest/pbmc3k.html#Clustering-the-neighborhood-graph

import numpy as np
import pandas as pd
import scanpy as sc
import anndata

adata=anndata.AnnData(features)
adata.obsm["X_tsne"] = tsne

sc.pp.neighbors(adata, n_neighbors=10, use_rep="X")
sc.tl.leiden(adata)

sc.pl.tsne(adata, color=["leiden"], save='_scanpy.pdf', show=False)

6. Citation - Shen et al. Miscell: An efficient self-supervised learning approach for dissecting single-cell transcriptome. iScience 2021(24):11.