Cells subjected to single cell RNA-seq have been through a lot, and they'd really just like to be alone now, please. If they cannot escape the cell social scene, you end up sequencing RNA from more than one cell to a barcode, creating a doublet when you expected single cell profiles. https://www.cell.com/cell-systems/fulltext/S2405-4712(20)30195-2
solo is a neural network framework to classify doublets, so that you can remove them from your data and clean your single cell profile.
We benchmarked solo against other doublet detection tools such as DoubletFinder and Scrublet, and found that it consistently outperformed them in terms of average precision. Additionally, Solo performed much better on a more complex tissue, mouse kidney.
Run the following to clone and set up ve.
git clone git@github.com:calico/solo.git && cd solo && conda create -n solo python=3.12 && conda activate solo && pip install -e .
Or install via pip
conda create -n solo python=3.12 && conda activate solo && pip install solo-sc
If you don't have conda follow the instructions here: https://docs.conda.io/projects/conda/en/latest/user-guide/install/
usage: solo [-h] -j MODEL_JSON_FILE -d DATA_PATH
[--set-reproducible-seed REPRODUCIBLE_SEED]
[--doublet-depth DOUBLET_DEPTH] [-g] [-a] [-o OUT_DIR]
[-r DOUBLET_RATIO] [-s SEED] [-e EXPECTED_NUMBER_OF_DOUBLETS] [-p]
[-recalibrate_scores] [--version] [--lr_st] [--lr_vae]
optional arguments:
-h, --help show this help message and exit
-j MODEL_JSON_FILE json file to pass VAE parameters (default: None)
-d DATA_PATH path to h5ad, loom, or 10x mtx dir cell by genes
counts (default: None)
--set-reproducible-seed REPRODUCIBLE_SEED
Reproducible seed, give an int to set seed (default:
None)
--doublet-depth DOUBLET_DEPTH
Depth multiplier for a doublet relative to the average
of its constituents (default: 2.0)
-g Run on GPU (default: True)
-a output modified anndata object with solo scores Only
works for anndata (default: False)
-o OUT_DIR
-r DOUBLET_RATIO Ratio of doublets to true cells (default: 2)
-s SEED Path to previous solo output directory. Seed VAE
models with previously trained solo model. Directory
structure is assumed to be the same as solo output
directory structure. should at least have a vae.pt a
pickled object of vae weights and a latent.npy an
np.ndarray of the latents of your cells. (default:
None)
-e EXPECTED_NUMBER_OF_DOUBLETS
Experimentally expected number of doublets (default:
None)
-p Plot outputs for solo (default: False)
-recalibrate_scores Recalibrate doublet scores (not recommended anymore)
(default: False)
--version Get version of solo-sc (default: False)
--lr_st
Learning rate used for solo.train (default: 1e-3)
--lr_vae
Learning rate used for vae (default: 1e-3)
Warning: If you are going directly from cellranger 10x output you may want to manually inspect your data prior to running solo.
model_json example:
{
"n_hidden": 384,
"n_latent": 64,
"n_layers": 1,
"cl_hidden": 128,
"cl_layers": 1,
"dropout_rate": 0.2,
"lr_st": 1e-3,
"valid_pct": 0.10
}
The suggested learning rates work best in most settings, but in case a ValueError occurs, you might consider changing the learning rates to 1e-5
Outputs:
-
is_doublet.npynp boolean array, true if a cell is a doublet, differs frompreds.npyif-e expected_number_of_doubletsparameter was used -
vaescVI directory for vae -
classifier.ptscVI directory for classifier -
latent.npylatent embedding for each cell -
preds.npydoublet predictions -
softmax_scores.npyupdated softmax of doublet scores (see paper), same asno_update_softmax_scores.npynow -
no_update_softmax_scores.npyraw softmax of doublet scores -
logit_scores.npylogit of doublet scores -
real_cells_dist.pdfhistogram of distribution of doublet scores -
accuracy.pdfaccuracy plot test vs train -
train_v_test_dist.pdfdoublet scores of test vs train -
roc.pdfroc of test vs train -
softmax_scores_sim.npysee above but for simulated doublets -
logit_scores_sim.npysee above but for simulated doublets -
preds_sim.npysee above but for simulated doublets -
is_doublet_sim.npysee above but for simulated doublets
Demultiplexing takes as input an h5ad file with only hashing counts. Counts can be obtained from your fastqs by using kite. See tutorial here: https://github.com/pachterlab/kite
usage: hashsolo [-h] [-j MODEL_JSON_FILE] [-o OUT_DIR] [-c CLUSTERING_DATA]
[-p PRE_EXISTING_CLUSTERS] [-q PLOT_NAME]
[-n NUMBER_OF_NOISE_BARCODES]
data_file
positional arguments:
data_file h5ad file containing cell hashing counts
optional arguments:
-h, --help show this help message and exit
-j MODEL_JSON_FILE json file to pass optional arguments (default: None)
-o OUT_DIR Output directory for results (default:
hashsolo_output)
-c CLUSTERING_DATA h5ad file with count transcriptional data to perform
clustering on (default: None)
-p PRE_EXISTING_CLUSTERS
column in cell_hashing_data_file.obs to specifying
different cell types or clusters (default: None)
-q PLOT_NAME name of plot to output (default: hashing_qc_plots.pdf)
-n NUMBER_OF_NOISE_BARCODES
Number of barcodes to use to create noise distribution
(default: None)
model_json example:
{
"priors": [0.01, 0.5, 0.49]
}
Priors is a list of the probability of the three hypotheses, negative, singlet,
or doublet that we test when demultiplexing cell hashing data. A negative cell's barcodes
doesn't have enough signal to identify its sample of origin. A singlet has
enough signal from single hashing barcode to associate the cell with ins
originating sample. A doublet is a cell barcode which has signal for more than one hashing barcode.
Depending on how you processed your cell hashing matrix before hand you may
want to set different priors. Under the assumption that you have subset your cell
barcodes using typical QC on your cell by genes matrix, e.g. min UMI counts,
percent mitochondrial reads, etc. We found the above setting of prior performed
well (see paper). If you have only done relatively light QC in transcriptome space
I'd suggest an even prior, e.g. [1./3., 1./3., 1./3.].
Outputs:
hashsoloed.h5adanndata with demultiplexing information in .obshashing_qc_plots.pngplots of probabilites for each cell
>>> from solo import hashsolo
>>> import anndata
>>> cell_hashing_data = anndata.read("cell_hashing_counts.h5ad")
>>> hashsolo.hashsolo(cell_hashing_data)
>>> cell_hashing_data.obs.head()
most_likeli_hypothesis cluster_feature negative_hypothesis_probability singlet_hypothesis_probability doublet_hypothesis_probability Classification
index
CCTTTCTGTCCGAACC 2 0 1.203673e-16 0.000002 0.999998 Doublet
CTGATAGGTGACTCAT 1 0 1.370633e-09 0.999920 0.000080 BatchF-GTGTGACGTATT_x
AGCTCTCGTTGTCTTT 1 0 2.369380e-13 0.996992 0.003008 BatchE-GAGGCTGAGCTA_x
GTGCGGTAGCGATGAC 1 0 1.579405e-09 0.999879 0.000121 BatchB-ACATGTTACCGT_x
AAATGCCTCTAACCGA 1 0 1.867626e-13 0.999707 0.000293 BatchB-ACATGTTACCGT_x
>>> demultiplex.plot_qc_checks_cell_hashing(cell_hashing_data)
most_likeli_hypothesis0 == Negative, 1 == Singlet, 2 == Doubletcluster_featurehow the cell hashing data was divided if specified or done automatically by giving a cell by genes anndata object to thecluster_dataargument when callingdemultiplex_cell_hashingnegative_hypothesis_probabilitysinglet_hypothesis_probabilitydoublet_hypothesis_probabilityClassificationThe sample of origin for the cell or whether it was a negative or doublet cell.