cbravo93
commented
2 years ago Hi @debbie28!
This is a great question! I will divide it in two parts:
-
'I have a cisTopic analysis in R, can I still use it for SCENIC+?'
Yes, you can! There is no need to repeat the topic modelling in Python :).
For this, you will need to do the following in R first:
# Export data to python library(arrow) path <- OUTDIR cisTopic_obj <- readRDS(PATH_TO_YOUR_CISTOPICOBJECT) modelMat <- modelMatSelection(cisTopic_obj, 'cell', 'Probability') modelMat <- as.data.frame(modelMat) write_feather(modelMat, sink=paste0(path, 'model_to_pycisTopic/cell_topic.feather')) modelMat <- modelMatSelection(cisTopic_obj, 'region', 'Probability', all.regions=TRUE) modelMat <- as.data.frame(modelMat) write_feather(modelMat, sink=paste0(path, 'model_to_pycisTopic/topic_region.feather')) # You can also save the count.matrix ct <- cisTopicObject@count.matrix ct <- as.data.frame(ct) write_feather(ct, sink=paste0(path, 'model_to_pycisTopic/count_matrix.feather'))And then in Python:
import pycisTopic ## 1. Initialize cisTopic object from pycisTopic.cistopic_class import * # Load count matrix matrix_path=PATH_TO_FRAGMENTS_MATRIX fragment_matrix = pd.read_feather(matrix_path) cisTopic_obj = create_cistopic_object(fragment_matrix) # Also add the cell annotation, cell_data should be a pandas df with cells as rows (cell names as index) and variables as columns cisTopic_obj.add_cell_data(cell_data) # 2. Add model from pycisTopic.utils import * model = load_cisTopic_model(PATH_TO_THE_FOLDER_WHERE_YOU_SAVED_DATA_FROM_R) cistopic_obj.add_LDA_model(model) # 3. Get imputed accessibility matrix from pycisTopic.diff_features import * imputed_acc_obj = impute_accessibility(cistopic_obj, selected_cells=None, selected_regions=None, scale_factor=10**6) # 4. DARs normalized_imputed_acc_obj = normalize_scores(imputed_acc_obj, scale_factor=10**4) variable_regions = find_highly_variable_features(normalized_imputed_acc_obj, min_disp = 0.05, min_mean = 0.0125, max_mean = 3, max_disp = np.inf, n_bins=20, n_top_features=None, plot=True) markers_dict= find_diff_features(cistopic_obj, imputed_acc_obj, variable='cell_type', # Here put the column of cell_data that you want to have DARs for var_features=variable_regions, contrasts=None, adjpval_thr=0.05, log2fc_thr=np.log2(1.5), n_cpu=5) # You can continue with pycistarget.Feel free to change the function if you do not want to work in feather format.
def load_cisTopic_model(path_to_cisTopic_model_matrices): metrics = None coherence = None marg_topic = None topic_ass = None cell_topic = pd.read_feather(path_to_cisTopic_model_matrices + "cell_topic.feather") cell_topic.index = ["Topic" + str(x) for x in range(1, cell_topic.shape[0] + 1)] topic_region = pd.read_feather( path_to_cisTopic_model_matrices + "topic_region.feather" ) topic_region.index = ["Topic" + str(x) for x in range(1, topic_region.shape[0] + 1)] topic_region = topic_region.T parameters = None model = cisTopicLDAModel( metrics, coherence, marg_topic, topic_ass, cell_topic, topic_region, parameters ) return model -
'I have an analysis with another tool (e.g. Signac/ArchR), can I still use it for SCENIC+?'
Here you have two options:
- Start from the count matrix: If you have performed QC steps and generated you fragment count matrix with another tool, you can directly start with the creation of the cisTopic object. In this tutorial you have an example on how to do this.
- You want to skip the topic modelling altogether: I would not recommend this option because we have not tested it, but technically it is possible. You would need to get 1) a denoised fragment matrix and 2) region sets. The region sets will be used as input for cisTarget. The region sets could be Differentially Accessible Regions (DARs); however we do recommend to use topics as well as they can capture regions with accessibility patterns without the need of specifying contrasts. If only using DARs you will have to make sure to make the right contrats (for example, all cell types versus the rest; higher cell type groups versus rest, etc). For example, in the cortex data set the topics reveal a topic of regions shared across all neurons; this pattern would be lost if you only do contrasts per cell type and not higher groups. In conclusion, it is possible to plug it in code-wise; but we can't assure results will be as good as the workflow we propose.
I will add these responses in a FAQs tab in the readthedocs page as I find them very relevant, thanks for the heads up!
C
georginafp
SeppeDeWinter
Hello! Thanks for developing SCENIC+. I really want to use this package to redo some of my multiome analyses. However I have already done a fair bit of work using Seurat/Signac and CisTopic in R, prior to this package coming out. I was wondering if I can use the CisTopic object I generated through R and use that as the input for SCENIC+ . I have done the TOPIC analysis and most of the initial preprocessing in R and it would be great if I could just import that to Python . Please let me know if this is possible.
Thanks Debbie