Empty factors - Spatial Transcripomics data

[rockdeme]

Hey!

I'm running MEFISTO on a 10X Visium sample based on the spatial transcriptomics tutorial. I have encountered an issue where some of the factors estimated by MEFISTO are all zeros, I'm not sure if this is an expect behaviour or something else I missed. I'm running MEFISTO with quite a lot (28) factors, maybe that's related to the problem.

Here is the code I'm running:

adata = sc.read_h5ad(adp)
sc.pp.filter_genes(adata, min_cells=len(adata) * 0.05)
sc.pp.normalize_total(adata, inplace=True)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata, flavor="seurat", n_top_genes=2000)

adata.obs = pd.concat([adata.obs, pd.DataFrame(adata.obsm["spatial"], columns=["imagerow", "imagecol"], index=adata.obs_names), ], axis=1)

ent = entry_point()
ent.set_data_options(use_float32=True)
ent.set_data_from_anndata(adata, features_subset="highly_variable")

ent.set_model_options(factors=28)
ent.set_train_options(save_interrupted=True)
ent.set_train_options(seed=2021)

n_inducing = 1000

ent.set_covariates([adata.obsm["spatial"]], covariates_names=["imagerow", "imagecol"])
ent.set_smooth_options(sparseGP=True, frac_inducing=n_inducing/adata.n_obs, start_opt=10, opt_freq=10)

ent.build()
ent.run()
factor_df = m.get_factors(df=True)

This is the output:

Training logs:

        #########################################################
        ###           __  __  ____  ______                    ### 
        ###          |  \/  |/ __ \|  ____/\    _             ### 
        ###          | \  / | |  | | |__ /  \ _| |_           ### 
        ###          | |\/| | |  | |  __/ /\ \_   _|          ###
        ###          | |  | | |__| | | / ____ \|_|            ###
        ###          |_|  |_|\____/|_|/_/    \_\              ###
        ###                                                   ### 
        ######################################################### 
       
 
        
use_float32 set to True: replacing float64 arrays by float32 arrays to speed up computations...

Loaded view='rna' group='group1' with N=2261 samples and D=2000 features...


Model options:
- Automatic Relevance Determination prior on the factors: False
- Automatic Relevance Determination prior on the weights: True
- Spike-and-slab prior on the factors: False
- Spike-and-slab prior on the weights: True
Likelihoods:
- View 0 (rna): gaussian


Loaded 2 covariate(s) for each sample...


##
## sparseGP set to True: using sparse Gaussian Process to speed up the training procedure
##


######################################
## Training the model with seed 2021 ##
######################################


ELBO before training: -76965346.50 

Iteration 1: time=7.73, ELBO=-2587610.86, deltaELBO=74377735.641 (96.63795334%), Factors=28
Iteration 2: time=7.19, ELBO=-2437343.13, deltaELBO=150267.728 (0.19524076%), Factors=28
Iteration 3: time=7.20, ELBO=-2381010.32, deltaELBO=56332.812 (0.07319244%), Factors=28
Iteration 4: time=7.20, ELBO=-2370936.18, deltaELBO=10074.139 (0.01308919%), Factors=28
Iteration 5: time=7.18, ELBO=-2366720.97, deltaELBO=4215.207 (0.00547676%), Factors=28
Iteration 6: time=7.68, ELBO=-2363330.84, deltaELBO=3390.130 (0.00440475%), Factors=28
Iteration 7: time=7.27, ELBO=-2360469.15, deltaELBO=2861.693 (0.00371816%), Factors=28
Iteration 8: time=7.27, ELBO=-2358012.11, deltaELBO=2457.035 (0.00319239%), Factors=28
Iteration 9: time=7.26, ELBO=-2355950.70, deltaELBO=2061.412 (0.00267836%), Factors=28
Optimising sigma node...
Iteration 10: time=277.38, ELBO=-2245553.37, deltaELBO=110397.331 (0.14343771%), Factors=28
Iteration 11: time=23.35, ELBO=-2324369.11, deltaELBO=-78815.744 (0.10240420%), Factors=28
Warning, lower bound is decreasing...�
Iteration 12: time=21.29, ELBO=-2315719.92, deltaELBO=8649.194 (0.01123778%), Factors=28
Iteration 13: time=22.62, ELBO=-2310927.10, deltaELBO=4792.816 (0.00622724%), Factors=28
Iteration 14: time=21.73, ELBO=-2307503.60, deltaELBO=3423.498 (0.00444810%), Factors=28
Iteration 15: time=22.93, ELBO=-2304850.19, deltaELBO=2653.415 (0.00344754%), Factors=28
Iteration 16: time=21.04, ELBO=-2302676.33, deltaELBO=2173.857 (0.00282446%), Factors=28
Iteration 17: time=23.12, ELBO=-2300786.11, deltaELBO=1890.225 (0.00245594%), Factors=28
Iteration 18: time=20.95, ELBO=-2299002.34, deltaELBO=1783.764 (0.00231762%), Factors=28
Iteration 19: time=22.73, ELBO=-2297239.95, deltaELBO=1762.397 (0.00228986%), Factors=28
Optimising sigma node...
Iteration 20: time=252.55, ELBO=-2295414.17, deltaELBO=1825.775 (0.00237220%), Factors=28
Iteration 21: time=23.38, ELBO=-2295356.47, deltaELBO=57.698 (0.00007497%), Factors=28
Iteration 22: time=21.31, ELBO=-2293432.14, deltaELBO=1924.338 (0.00250027%), Factors=28
Iteration 23: time=23.37, ELBO=-2291737.80, deltaELBO=1694.332 (0.00220142%), Factors=28
Iteration 24: time=21.26, ELBO=-2290347.46, deltaELBO=1390.345 (0.00180646%), Factors=28
Iteration 25: time=23.39, ELBO=-2289241.89, deltaELBO=1105.566 (0.00143645%), Factors=28
Iteration 26: time=21.33, ELBO=-2288333.65, deltaELBO=908.242 (0.00118007%), Factors=28
Iteration 27: time=23.49, ELBO=-2287549.41, deltaELBO=784.242 (0.00101896%), Factors=28
Iteration 28: time=21.88, ELBO=-2286852.86, deltaELBO=696.553 (0.00090502%), Factors=28
Iteration 29: time=25.29, ELBO=-2286229.91, deltaELBO=622.948 (0.00080939%), Factors=28
Optimising sigma node...
Iteration 30: time=359.29, ELBO=-2285490.63, deltaELBO=739.273 (0.00096053%), Factors=28
Iteration 31: time=27.38, ELBO=-2293869.65, deltaELBO=-8379.011 (0.01088673%), Factors=28
Warning, lower bound is decreasing...�
Iteration 32: time=24.32, ELBO=-2292963.65, deltaELBO=905.998 (0.00117715%), Factors=28
Iteration 33: time=28.33, ELBO=-2292207.38, deltaELBO=756.265 (0.00098260%), Factors=28
Iteration 34: time=25.16, ELBO=-2291545.92, deltaELBO=661.464 (0.00085943%), Factors=28
Iteration 35: time=26.70, ELBO=-2290951.27, deltaELBO=594.646 (0.00077262%), Factors=28
Iteration 36: time=24.08, ELBO=-2290412.84, deltaELBO=538.431 (0.00069958%), Factors=28
Iteration 37: time=31.96, ELBO=-2289917.60, deltaELBO=495.242 (0.00064346%), Factors=28
Iteration 38: time=25.61, ELBO=-2289461.21, deltaELBO=456.386 (0.00059298%), Factors=28
Iteration 39: time=24.18, ELBO=-2289038.30, deltaELBO=422.916 (0.00054949%), Factors=28
Optimising sigma node...
Iteration 40: time=255.44, ELBO=-2288558.93, deltaELBO=479.362 (0.00062283%), Factors=28
Iteration 41: time=23.36, ELBO=-2289573.09, deltaELBO=-1014.154 (0.00131768%), Factors=28
Warning, lower bound is decreasing...�
Iteration 42: time=21.33, ELBO=-2289146.12, deltaELBO=426.972 (0.00055476%), Factors=28
Iteration 43: time=23.29, ELBO=-2288793.99, deltaELBO=352.124 (0.00045751%), Factors=28
Iteration 44: time=21.20, ELBO=-2288474.15, deltaELBO=319.842 (0.00041557%), Factors=28

Converged!



#######################
## Training finished ##
#######################