Spatial molecular niche discovery

Here we will use seqFISH Mouse Organogenesis datasets as an example to illustrate how Garfield performs spatial molecular niche discovery for multiple batches. Data access: https://marionilab.cruk.cam.ac.uk/SpatialMouseAtlas/.

[1]:

import os
import pandas as pd
import numpy as np

os.chdir('/data2/zhouwg_data/project/Garfield')
os.getcwd()

[1]:

'/data2/zhouwg_data/project/Garfield'

[2]:

# load packages
import os
import warnings
import Garfield as gf
import scanpy as sc
from mudata import MuData
warnings.simplefilter(action="ignore", category=FutureWarning)
warnings.simplefilter(action='ignore', category=UserWarning)

gf.__version__

[2]:

'1.0.0'

[3]:

# Load data
# Define paths
st_data_folder_path = "/home/zhouweige/zhouwg_data/project/Garfield_tutorials/data" # spatially resolved transcriptomics data
dataset = "seqfish_mouse_organogenesis"
cell_type_key = "celltype_mapped_refined"

adata_list = []
batch_indeces = list(np.arange(6) + 1)
batches = ["embryo1_z2",
           "embryo1_z5",
           "embryo2_z2",
           "embryo2_z5",
           "embryo3_z2",
           "embryo3_z5"]

for batch_idx, batch in zip(batch_indeces, batches):
    adata = sc.read_h5ad(f"{st_data_folder_path}/{dataset}_batch{batch_idx}.h5ad")
    adata.var_names_make_unique(join="++")
    # make spot name unique
    # adata.obs_names = [x+'_'+batch for x in adata.obs_names]
    adata_list.append(adata)

[4]:

import anndata as ad

adata_concat = ad.concat(adata_list, label="slice_name", keys=batches)
adata_concat

[4]:

AnnData object with n_obs × n_vars = 52568 × 351
    obs: 'Area', 'celltype_mapped_refined', 'sample', 'batch', 'slice_name'
    obsm: 'spatial'
    layers: 'counts'

[5]:

# check the batch information of this dataset
adata_concat.obs['batch'].value_counts()

[5]:

batch
embryo3_z2    11169
embryo1_z2    10150
embryo3_z5     9408
embryo2_z2     7661
embryo1_z5     7656
embryo2_z5     6524
Name: count, dtype: int64

[6]:

# Ensure adata.X is counts.
adata_concat.X = adata_concat.layers['counts'].copy()
adata_concat.X.max()

[6]:

141.0

Integrating spatially resolved transcriptomics data using Garfield

For spatial niche discovery from spatially resolved transcriptomics data, we should adjust the following paremeters, and all parameter introductions can be found in Garfield_Model_Parameters.

[7]:

# set workdir
workdir = '/home/zhouweige/zhouwg_data/project/Garfield_tutorials/result/garfield_spRNA_embryo'
gf.settings.set_workdir(workdir)

### modify parameter
user_config = dict(
    ## Input options
    adata_list=adata_concat,
    profile='spatial',
    data_type='single-modal',
    sample_col='batch',
    weight=0.5,
    ## Preprocessing options
    graph_const_method='mu_std', # mu_std, Radius, KNN, Squidpy
    used_hvg=True,
    min_cells=3,
    min_features=0,
    keep_mt=False,
    target_sum=1e4,
    rna_n_top_features=3000,
    n_components=50,
    n_neighbors=5,
    metric='euclidean',
    svd_solver='arpack',
    # datasets
    used_pca_feat=False,
    adj_key='connectivities',
    # data split parameters
    edge_val_ratio=0.1,
    edge_test_ratio=0.,
    node_val_ratio=0.1,
    node_test_ratio=0.,
    ## Model options
    augment_type='svd',
    svd_q=5,
    use_FCencoder=False,
    conv_type='GATv2Conv', # GAT or GATv2Conv or GCN
    gnn_layer=2,
    hidden_dims=[128, 128],
    bottle_neck_neurons=20,
    cluster_num=20,
    drop_feature_rate=0.2,
    drop_edge_rate=0.2,
    num_heads=3,
    dropout=0.2,
    concat=True,
    used_edge_weight=True,
    used_DSBN=False,
    used_mmd=False,
    # data loader parameters
    num_neighbors=5,
    loaders_n_hops=2,
    edge_batch_size=4096,
    node_batch_size=128, # None
    # loss parameters
    include_edge_recon_loss=True,
    include_gene_expr_recon_loss=True,
    lambda_latent_contrastive_instanceloss=1.0,
    lambda_latent_contrastive_clusterloss=0.5,
    lambda_gene_expr_recon=10.,
    lambda_edge_recon=5., #
    lambda_latent_adj_recon_loss=1.0,
    lambda_omics_recon_mmd_loss=0.2,
    # train parameters
    n_epochs_no_edge_recon=0,
    learning_rate=0.001,
    weight_decay=1e-05,
    gradient_clipping=5,
    # other parameters
    latent_key='garfield_latent',
    reload_best_model=True,
    use_early_stopping=True,
    early_stopping_kwargs=None,
    monitor=True,
    device_id=1,
    seed=2024,
    verbose=True
)
dict_config = gf.settings.set_gf_params(user_config)

Saving results in: /home/zhouweige/zhouwg_data/project/Garfield_tutorials/result/garfield_spRNA_embryo

[8]:

from Garfield.model import Garfield

# Initialize model
model = Garfield(dict_config)

--- DATA LOADING AND PREPROCESSING ---

/home/zhouweige/anaconda3/envs/Garfield/lib/python3.9/site-packages/scipy/sparse/_index.py:143: SparseEfficiencyWarning: Changing the sparsity structure of a csr_matrix is expensive. lil_matrix is more efficient.
  self._set_arrayXarray(i, j, x)

------Calculating spatial graph...
COSINE SIM GRAPH DECODER -> dropout_rate: 0.2

[9]:

# Train model
model.train()


--- INITIALIZING TRAINER ---
Using GPU: device-1
Number of training nodes: 47311
Number of validation nodes: 5257
Number of training edges: 272860
Number of validation edges: 30317
Edge batch size: 4096
Node batch size: 128

--- MODEL TRAINING ---
Epoch 1/100 |--------------------| 1.0% val_auroc_score: 0.8252; val_auprc_score: 0.8270; val_best_acc_score: 0.7531; val_best_f1_score: 0.7606; train_kl_reg_loss: 143.9980; train_edge_recon_loss: 5885.2862; train_gene_expr_recon_loss: 670891.8582; train_lambda_latent_adj_recon_loss: 2975.3924; train_lambda_latent_contrastive_instanceloss: 7.0683; train_lambda_latent_contrastive_clusterloss: 3.6366; train_global_loss: 679907.2332; train_optim_loss: 679907.2332; val_kl_reg_loss: 99.1661; val_edge_recon_loss: 5287.1518; val_gene_expr_recon_loss: 576232.7344; val_lambda_latent_adj_recon_loss: 2698.1509; val_lambda_latent_contrastive_instanceloss: 6.4858; val_lambda_latent_contrastive_clusterloss: 3.4129; val_global_loss: 584327.1250; val_optim_loss: 584327.1250
Epoch 2/100 |--------------------| 2.0% val_auroc_score: 0.8855; val_auprc_score: 0.8968; val_best_acc_score: 0.8180; val_best_f1_score: 0.8165; train_kl_reg_loss: 162.3178; train_edge_recon_loss: 5570.9410; train_gene_expr_recon_loss: 581223.2808; train_lambda_latent_adj_recon_loss: 2754.0141; train_lambda_latent_contrastive_instanceloss: 6.6238; train_lambda_latent_contrastive_clusterloss: 3.3036; train_global_loss: 589720.4795; train_optim_loss: 589720.4795; val_kl_reg_loss: 179.9585; val_edge_recon_loss: 5172.7614; val_gene_expr_recon_loss: 540662.9297; val_lambda_latent_adj_recon_loss: 3102.7676; val_lambda_latent_contrastive_instanceloss: 6.3932; val_lambda_latent_contrastive_clusterloss: 2.9511; val_global_loss: 549127.7500; val_optim_loss: 549127.7500
Epoch 3/100 |--------------------| 3.0% val_auroc_score: 0.8980; val_auprc_score: 0.9121; val_best_acc_score: 0.8374; val_best_f1_score: 0.8321; train_kl_reg_loss: 186.9358; train_edge_recon_loss: 5424.3385; train_gene_expr_recon_loss: 560887.7957; train_lambda_latent_adj_recon_loss: 2676.9108; train_lambda_latent_contrastive_instanceloss: 6.4969; train_lambda_latent_contrastive_clusterloss: 2.9072; train_global_loss: 569185.3862; train_optim_loss: 569185.3862; val_kl_reg_loss: 154.7024; val_edge_recon_loss: 4827.5549; val_gene_expr_recon_loss: 533907.7031; val_lambda_latent_adj_recon_loss: 2409.7722; val_lambda_latent_contrastive_instanceloss: 6.3748; val_lambda_latent_contrastive_clusterloss: 2.6642; val_global_loss: 541308.7891; val_optim_loss: 541308.7891
Epoch 4/100 |--------------------| 4.0% val_auroc_score: 0.8965; val_auprc_score: 0.9150; val_best_acc_score: 0.8383; val_best_f1_score: 0.8313; train_kl_reg_loss: 192.4402; train_edge_recon_loss: 5259.7526; train_gene_expr_recon_loss: 553952.8601; train_lambda_latent_adj_recon_loss: 2243.6234; train_lambda_latent_contrastive_instanceloss: 6.4629; train_lambda_latent_contrastive_clusterloss: 2.6861; train_global_loss: 561657.8265; train_optim_loss: 561657.8265; val_kl_reg_loss: 166.0171; val_edge_recon_loss: 4679.4329; val_gene_expr_recon_loss: 524051.4688; val_lambda_latent_adj_recon_loss: 1909.6371; val_lambda_latent_contrastive_instanceloss: 6.3632; val_lambda_latent_contrastive_clusterloss: 2.5104; val_global_loss: 530815.4414; val_optim_loss: 530815.4414
Epoch 5/100 |█-------------------| 5.0% val_auroc_score: 0.8926; val_auprc_score: 0.9119; val_best_acc_score: 0.8347; val_best_f1_score: 0.8258; train_kl_reg_loss: 212.0395; train_edge_recon_loss: 5125.5335; train_gene_expr_recon_loss: 546801.5289; train_lambda_latent_adj_recon_loss: 1847.5145; train_lambda_latent_contrastive_instanceloss: 6.4340; train_lambda_latent_contrastive_clusterloss: 2.5556; train_global_loss: 553995.6110; train_optim_loss: 553995.6110; val_kl_reg_loss: 201.1441; val_edge_recon_loss: 4601.6688; val_gene_expr_recon_loss: 521174.4688; val_lambda_latent_adj_recon_loss: 1845.8800; val_lambda_latent_contrastive_instanceloss: 6.3566; val_lambda_latent_contrastive_clusterloss: 2.4339; val_global_loss: 527831.9531; val_optim_loss: 527831.9531
Epoch 6/100 |█-------------------| 6.0% val_auroc_score: 0.8901; val_auprc_score: 0.9112; val_best_acc_score: 0.8348; val_best_f1_score: 0.8245; train_kl_reg_loss: 242.8385; train_edge_recon_loss: 5071.4030; train_gene_expr_recon_loss: 542545.4767; train_lambda_latent_adj_recon_loss: 1764.0965; train_lambda_latent_contrastive_instanceloss: 6.4185; train_lambda_latent_contrastive_clusterloss: 2.4657; train_global_loss: 549632.7099; train_optim_loss: 549632.7099; val_kl_reg_loss: 213.3091; val_edge_recon_loss: 4568.9492; val_gene_expr_recon_loss: 520497.9961; val_lambda_latent_adj_recon_loss: 1773.7744; val_lambda_latent_contrastive_instanceloss: 6.3481; val_lambda_latent_contrastive_clusterloss: 2.3450; val_global_loss: 527062.6992; val_optim_loss: 527062.6992
Epoch 7/100 |█-------------------| 7.0% val_auroc_score: 0.8930; val_auprc_score: 0.9151; val_best_acc_score: 0.8371; val_best_f1_score: 0.8286; train_kl_reg_loss: 256.7653; train_edge_recon_loss: 5031.2601; train_gene_expr_recon_loss: 540507.4622; train_lambda_latent_adj_recon_loss: 1670.0192; train_lambda_latent_contrastive_instanceloss: 6.4079; train_lambda_latent_contrastive_clusterloss: 2.3941; train_global_loss: 547474.3088; train_optim_loss: 547474.3088; val_kl_reg_loss: 240.0973; val_edge_recon_loss: 4518.9738; val_gene_expr_recon_loss: 514080.4648; val_lambda_latent_adj_recon_loss: 1721.5623; val_lambda_latent_contrastive_instanceloss: 6.3438; val_lambda_latent_contrastive_clusterloss: 2.2874; val_global_loss: 520569.7266; val_optim_loss: 520569.7266
Epoch 8/100 |█-------------------| 8.0% val_auroc_score: 0.8949; val_auprc_score: 0.9146; val_best_acc_score: 0.8361; val_best_f1_score: 0.8278; train_kl_reg_loss: 275.7386; train_edge_recon_loss: 5009.1849; train_gene_expr_recon_loss: 537798.3410; train_lambda_latent_adj_recon_loss: 1635.0023; train_lambda_latent_contrastive_instanceloss: 6.3948; train_lambda_latent_contrastive_clusterloss: 2.3364; train_global_loss: 544726.9841; train_optim_loss: 544726.9841; val_kl_reg_loss: 231.0863; val_edge_recon_loss: 4497.0161; val_gene_expr_recon_loss: 514525.2461; val_lambda_latent_adj_recon_loss: 1661.7053; val_lambda_latent_contrastive_instanceloss: 6.3447; val_lambda_latent_contrastive_clusterloss: 2.2413; val_global_loss: 520923.6406; val_optim_loss: 520923.6406
Epoch 9/100 |█-------------------| 9.0% val_auroc_score: 0.8962; val_auprc_score: 0.9164; val_best_acc_score: 0.8379; val_best_f1_score: 0.8299; train_kl_reg_loss: 279.5599; train_edge_recon_loss: 4994.0874; train_gene_expr_recon_loss: 532982.8526; train_lambda_latent_adj_recon_loss: 1589.0474; train_lambda_latent_contrastive_instanceloss: 6.3902; train_lambda_latent_contrastive_clusterloss: 2.3029; train_global_loss: 539854.2388; train_optim_loss: 539854.2388; val_kl_reg_loss: 231.6465; val_edge_recon_loss: 4475.7152; val_gene_expr_recon_loss: 510907.1016; val_lambda_latent_adj_recon_loss: 1548.3639; val_lambda_latent_contrastive_instanceloss: 6.3354; val_lambda_latent_contrastive_clusterloss: 2.2099; val_global_loss: 517171.3984; val_optim_loss: 517171.3984
Epoch 10/100 |██------------------| 10.0% val_auroc_score: 0.8931; val_auprc_score: 0.9150; val_best_acc_score: 0.8359; val_best_f1_score: 0.8270; train_kl_reg_loss: 292.8605; train_edge_recon_loss: 4982.7499; train_gene_expr_recon_loss: 532560.5868; train_lambda_latent_adj_recon_loss: 1577.1255; train_lambda_latent_contrastive_instanceloss: 6.3810; train_lambda_latent_contrastive_clusterloss: 2.2637; train_global_loss: 539421.9669; train_optim_loss: 539421.9669; val_kl_reg_loss: 248.0701; val_edge_recon_loss: 4458.2536; val_gene_expr_recon_loss: 510130.5156; val_lambda_latent_adj_recon_loss: 1589.7175; val_lambda_latent_contrastive_instanceloss: 6.3374; val_lambda_latent_contrastive_clusterloss: 2.1822; val_global_loss: 516435.0703; val_optim_loss: 516435.0703
Epoch 11/100 |██------------------| 11.0% val_auroc_score: 0.8961; val_auprc_score: 0.9165; val_best_acc_score: 0.8388; val_best_f1_score: 0.8299; train_kl_reg_loss: 295.9002; train_edge_recon_loss: 4969.2420; train_gene_expr_recon_loss: 529195.2411; train_lambda_latent_adj_recon_loss: 1546.5997; train_lambda_latent_contrastive_instanceloss: 6.3753; train_lambda_latent_contrastive_clusterloss: 2.2393; train_global_loss: 536015.5998; train_optim_loss: 536015.5998; val_kl_reg_loss: 244.7273; val_edge_recon_loss: 4457.2879; val_gene_expr_recon_loss: 512043.1133; val_lambda_latent_adj_recon_loss: 1519.4812; val_lambda_latent_contrastive_instanceloss: 6.3323; val_lambda_latent_contrastive_clusterloss: 2.1609; val_global_loss: 518273.0977; val_optim_loss: 518273.0977
Epoch 12/100 |██------------------| 12.0% val_auroc_score: 0.8965; val_auprc_score: 0.9170; val_best_acc_score: 0.8384; val_best_f1_score: 0.8299; train_kl_reg_loss: 290.2068; train_edge_recon_loss: 4958.8105; train_gene_expr_recon_loss: 527256.4380; train_lambda_latent_adj_recon_loss: 1502.3812; train_lambda_latent_contrastive_instanceloss: 6.3720; train_lambda_latent_contrastive_clusterloss: 2.2209; train_global_loss: 534016.4272; train_optim_loss: 534016.4272; val_kl_reg_loss: 252.6359; val_edge_recon_loss: 4451.9817; val_gene_expr_recon_loss: 508041.3281; val_lambda_latent_adj_recon_loss: 1524.9273; val_lambda_latent_contrastive_instanceloss: 6.3239; val_lambda_latent_contrastive_clusterloss: 2.1251; val_global_loss: 514279.3203; val_optim_loss: 514279.3203
Epoch 13/100 |██------------------| 13.0% val_auroc_score: 0.8998; val_auprc_score: 0.9198; val_best_acc_score: 0.8424; val_best_f1_score: 0.8330; train_kl_reg_loss: 301.1507; train_edge_recon_loss: 4951.7350; train_gene_expr_recon_loss: 526170.7831; train_lambda_latent_adj_recon_loss: 1523.9489; train_lambda_latent_contrastive_instanceloss: 6.3681; train_lambda_latent_contrastive_clusterloss: 2.2055; train_global_loss: 532956.1959; train_optim_loss: 532956.1959; val_kl_reg_loss: 244.9166; val_edge_recon_loss: 4446.4363; val_gene_expr_recon_loss: 508866.7656; val_lambda_latent_adj_recon_loss: 1513.7197; val_lambda_latent_contrastive_instanceloss: 6.3230; val_lambda_latent_contrastive_clusterloss: 2.1247; val_global_loss: 515080.2930; val_optim_loss: 515080.2930
Epoch 14/100 |██------------------| 14.0% val_auroc_score: 0.8963; val_auprc_score: 0.9178; val_best_acc_score: 0.8417; val_best_f1_score: 0.8318; train_kl_reg_loss: 291.3508; train_edge_recon_loss: 4951.6723; train_gene_expr_recon_loss: 525927.7659; train_lambda_latent_adj_recon_loss: 1480.0718; train_lambda_latent_contrastive_instanceloss: 6.3687; train_lambda_latent_contrastive_clusterloss: 2.1943; train_global_loss: 532659.4361; train_optim_loss: 532659.4361; val_kl_reg_loss: 252.7811; val_edge_recon_loss: 4446.4618; val_gene_expr_recon_loss: 503508.1758; val_lambda_latent_adj_recon_loss: 1509.3852; val_lambda_latent_contrastive_instanceloss: 6.3274; val_lambda_latent_contrastive_clusterloss: 2.1220; val_global_loss: 509725.2578; val_optim_loss: 509725.2578
Epoch 15/100 |███-----------------| 15.0% val_auroc_score: 0.9002; val_auprc_score: 0.9203; val_best_acc_score: 0.8422; val_best_f1_score: 0.8337; train_kl_reg_loss: 285.8749; train_edge_recon_loss: 4943.2590; train_gene_expr_recon_loss: 525165.6632; train_lambda_latent_adj_recon_loss: 1443.0932; train_lambda_latent_contrastive_instanceloss: 6.3647; train_lambda_latent_contrastive_clusterloss: 2.1833; train_global_loss: 531846.4496; train_optim_loss: 531846.4496; val_kl_reg_loss: 224.6013; val_edge_recon_loss: 4416.8882; val_gene_expr_recon_loss: 505968.2500; val_lambda_latent_adj_recon_loss: 1393.3304; val_lambda_latent_contrastive_instanceloss: 6.3290; val_lambda_latent_contrastive_clusterloss: 2.1134; val_global_loss: 512011.5117; val_optim_loss: 512011.5117
Epoch 16/100 |███-----------------| 16.0% val_auroc_score: 0.9030; val_auprc_score: 0.9240; val_best_acc_score: 0.8481; val_best_f1_score: 0.8384; train_kl_reg_loss: 285.2293; train_edge_recon_loss: 4943.5495; train_gene_expr_recon_loss: 524888.5089; train_lambda_latent_adj_recon_loss: 1435.2312; train_lambda_latent_contrastive_instanceloss: 6.3615; train_lambda_latent_contrastive_clusterloss: 2.1716; train_global_loss: 531561.0555; train_optim_loss: 531561.0555; val_kl_reg_loss: 236.5831; val_edge_recon_loss: 4435.7971; val_gene_expr_recon_loss: 511203.5156; val_lambda_latent_adj_recon_loss: 1454.1355; val_lambda_latent_contrastive_instanceloss: 6.3278; val_lambda_latent_contrastive_clusterloss: 2.1142; val_global_loss: 517338.4727; val_optim_loss: 517338.4727
Epoch 17/100 |███-----------------| 17.0% val_auroc_score: 0.9034; val_auprc_score: 0.9230; val_best_acc_score: 0.8467; val_best_f1_score: 0.8380; train_kl_reg_loss: 295.6101; train_edge_recon_loss: 4942.3230; train_gene_expr_recon_loss: 522807.2085; train_lambda_latent_adj_recon_loss: 1461.6460; train_lambda_latent_contrastive_instanceloss: 6.3606; train_lambda_latent_contrastive_clusterloss: 2.1564; train_global_loss: 529515.3074; train_optim_loss: 529515.3074; val_kl_reg_loss: 247.3454; val_edge_recon_loss: 4424.4774; val_gene_expr_recon_loss: 501113.3594; val_lambda_latent_adj_recon_loss: 1437.2095; val_lambda_latent_contrastive_instanceloss: 6.3199; val_lambda_latent_contrastive_clusterloss: 2.1066; val_global_loss: 507230.8242; val_optim_loss: 507230.8242
Epoch 18/100 |███-----------------| 18.0% val_auroc_score: 0.9023; val_auprc_score: 0.9220; val_best_acc_score: 0.8428; val_best_f1_score: 0.8346; train_kl_reg_loss: 296.2562; train_edge_recon_loss: 4928.3164; train_gene_expr_recon_loss: 523123.4981; train_lambda_latent_adj_recon_loss: 1441.5480; train_lambda_latent_contrastive_instanceloss: 6.3559; train_lambda_latent_contrastive_clusterloss: 2.1503; train_global_loss: 529798.1283; train_optim_loss: 529798.1283; val_kl_reg_loss: 252.3596; val_edge_recon_loss: 4441.3446; val_gene_expr_recon_loss: 508097.1133; val_lambda_latent_adj_recon_loss: 1453.0274; val_lambda_latent_contrastive_instanceloss: 6.3211; val_lambda_latent_contrastive_clusterloss: 2.0710; val_global_loss: 514252.2422; val_optim_loss: 514252.2422
Epoch 19/100 |███-----------------| 19.0% val_auroc_score: 0.9030; val_auprc_score: 0.9215; val_best_acc_score: 0.8447; val_best_f1_score: 0.8360; train_kl_reg_loss: 303.9804; train_edge_recon_loss: 4928.4915; train_gene_expr_recon_loss: 521529.1665; train_lambda_latent_adj_recon_loss: 1439.9777; train_lambda_latent_contrastive_instanceloss: 6.3528; train_lambda_latent_contrastive_clusterloss: 2.1325; train_global_loss: 528210.1129; train_optim_loss: 528210.1129; val_kl_reg_loss: 251.3782; val_edge_recon_loss: 4421.6002; val_gene_expr_recon_loss: 502723.9062; val_lambda_latent_adj_recon_loss: 1397.3658; val_lambda_latent_contrastive_instanceloss: 6.3165; val_lambda_latent_contrastive_clusterloss: 2.0860; val_global_loss: 508802.6523; val_optim_loss: 508802.6523
Epoch 20/100 |████----------------| 20.0% val_auroc_score: 0.9049; val_auprc_score: 0.9237; val_best_acc_score: 0.8466; val_best_f1_score: 0.8377; train_kl_reg_loss: 298.4227; train_edge_recon_loss: 4924.6892; train_gene_expr_recon_loss: 521198.7281; train_lambda_latent_adj_recon_loss: 1411.4458; train_lambda_latent_contrastive_instanceloss: 6.3539; train_lambda_latent_contrastive_clusterloss: 2.1450; train_global_loss: 527841.7854; train_optim_loss: 527841.7854; val_kl_reg_loss: 242.6093; val_edge_recon_loss: 4420.7876; val_gene_expr_recon_loss: 509210.9922; val_lambda_latent_adj_recon_loss: 1398.2573; val_lambda_latent_contrastive_instanceloss: 6.3150; val_lambda_latent_contrastive_clusterloss: 2.0734; val_global_loss: 515281.0430; val_optim_loss: 515281.0430
Epoch 21/100 |████----------------| 21.0% val_auroc_score: 0.9078; val_auprc_score: 0.9263; val_best_acc_score: 0.8495; val_best_f1_score: 0.8414; train_kl_reg_loss: 298.6156; train_edge_recon_loss: 4921.0351; train_gene_expr_recon_loss: 520087.0401; train_lambda_latent_adj_recon_loss: 1415.8348; train_lambda_latent_contrastive_instanceloss: 6.3515; train_lambda_latent_contrastive_clusterloss: 2.1229; train_global_loss: 526731.0033; train_optim_loss: 526731.0033; val_kl_reg_loss: 244.2098; val_edge_recon_loss: 4419.9619; val_gene_expr_recon_loss: 505643.0352; val_lambda_latent_adj_recon_loss: 1373.9524; val_lambda_latent_contrastive_instanceloss: 6.3196; val_lambda_latent_contrastive_clusterloss: 2.0759; val_global_loss: 511689.5391; val_optim_loss: 511689.5391

Reducing learning rate: metric has not improved more than 0.0 in the last 4 epochs.
New learning rate is 0.0001.

Epoch 22/100 |████----------------| 22.0% val_auroc_score: 0.9079; val_auprc_score: 0.9268; val_best_acc_score: 0.8495; val_best_f1_score: 0.8414; train_kl_reg_loss: 296.6114; train_edge_recon_loss: 4913.8968; train_gene_expr_recon_loss: 517173.4622; train_lambda_latent_adj_recon_loss: 1373.3456; train_lambda_latent_contrastive_instanceloss: 6.3477; train_lambda_latent_contrastive_clusterloss: 2.1230; train_global_loss: 523765.7906; train_optim_loss: 523765.7906; val_kl_reg_loss: 233.6331; val_edge_recon_loss: 4394.0042; val_gene_expr_recon_loss: 502743.5547; val_lambda_latent_adj_recon_loss: 1336.0905; val_lambda_latent_contrastive_instanceloss: 6.3144; val_lambda_latent_contrastive_clusterloss: 2.0668; val_global_loss: 508715.6602; val_optim_loss: 508715.6602
Epoch 23/100 |████----------------| 23.0% val_auroc_score: 0.9069; val_auprc_score: 0.9259; val_best_acc_score: 0.8483; val_best_f1_score: 0.8396; train_kl_reg_loss: 301.2392; train_edge_recon_loss: 4906.6374; train_gene_expr_recon_loss: 516599.6287; train_lambda_latent_adj_recon_loss: 1386.9922; train_lambda_latent_contrastive_instanceloss: 6.3471; train_lambda_latent_contrastive_clusterloss: 2.1182; train_global_loss: 523202.9646; train_optim_loss: 523202.9646; val_kl_reg_loss: 242.8831; val_edge_recon_loss: 4404.0029; val_gene_expr_recon_loss: 503694.2695; val_lambda_latent_adj_recon_loss: 1355.2806; val_lambda_latent_contrastive_instanceloss: 6.3104; val_lambda_latent_contrastive_clusterloss: 2.0523; val_global_loss: 509704.8086; val_optim_loss: 509704.8086
Epoch 24/100 |████----------------| 24.0% val_auroc_score: 0.9086; val_auprc_score: 0.9267; val_best_acc_score: 0.8502; val_best_f1_score: 0.8420; train_kl_reg_loss: 306.9987; train_edge_recon_loss: 4906.5948; train_gene_expr_recon_loss: 515721.0051; train_lambda_latent_adj_recon_loss: 1385.8733; train_lambda_latent_contrastive_instanceloss: 6.3446; train_lambda_latent_contrastive_clusterloss: 2.1204; train_global_loss: 522328.9398; train_optim_loss: 522328.9398; val_kl_reg_loss: 250.2436; val_edge_recon_loss: 4398.5364; val_gene_expr_recon_loss: 503862.2031; val_lambda_latent_adj_recon_loss: 1382.0446; val_lambda_latent_contrastive_instanceloss: 6.3199; val_lambda_latent_contrastive_clusterloss: 2.0455; val_global_loss: 509901.3828; val_optim_loss: 509901.3828
Epoch 25/100 |█████---------------| 25.0% val_auroc_score: 0.9080; val_auprc_score: 0.9265; val_best_acc_score: 0.8482; val_best_f1_score: 0.8414; train_kl_reg_loss: 314.3502; train_edge_recon_loss: 4903.8369; train_gene_expr_recon_loss: 517129.7495; train_lambda_latent_adj_recon_loss: 1401.8595; train_lambda_latent_contrastive_instanceloss: 6.3460; train_lambda_latent_contrastive_clusterloss: 2.1140; train_global_loss: 523758.2575; train_optim_loss: 523758.2575; val_kl_reg_loss: 252.1183; val_edge_recon_loss: 4395.6864; val_gene_expr_recon_loss: 498417.9414; val_lambda_latent_adj_recon_loss: 1355.1604; val_lambda_latent_contrastive_instanceloss: 6.3055; val_lambda_latent_contrastive_clusterloss: 2.0429; val_global_loss: 504429.2539; val_optim_loss: 504429.2539
Epoch 26/100 |█████---------------| 26.0% val_auroc_score: 0.9094; val_auprc_score: 0.9265; val_best_acc_score: 0.8508; val_best_f1_score: 0.8432; train_kl_reg_loss: 321.0471; train_edge_recon_loss: 4900.5190; train_gene_expr_recon_loss: 516130.0588; train_lambda_latent_adj_recon_loss: 1412.2932; train_lambda_latent_contrastive_instanceloss: 6.3432; train_lambda_latent_contrastive_clusterloss: 2.1062; train_global_loss: 522772.3657; train_optim_loss: 522772.3657; val_kl_reg_loss: 256.5296; val_edge_recon_loss: 4388.4486; val_gene_expr_recon_loss: 500430.6328; val_lambda_latent_adj_recon_loss: 1343.0289; val_lambda_latent_contrastive_instanceloss: 6.3153; val_lambda_latent_contrastive_clusterloss: 2.0634; val_global_loss: 506427.0117; val_optim_loss: 506427.0117
Epoch 27/100 |█████---------------| 27.0% val_auroc_score: 0.9099; val_auprc_score: 0.9281; val_best_acc_score: 0.8512; val_best_f1_score: 0.8438; train_kl_reg_loss: 322.1430; train_edge_recon_loss: 4898.4001; train_gene_expr_recon_loss: 515153.5536; train_lambda_latent_adj_recon_loss: 1404.5756; train_lambda_latent_contrastive_instanceloss: 6.3393; train_lambda_latent_contrastive_clusterloss: 2.1024; train_global_loss: 521787.1152; train_optim_loss: 521787.1152; val_kl_reg_loss: 257.4874; val_edge_recon_loss: 4391.9070; val_gene_expr_recon_loss: 505991.1016; val_lambda_latent_adj_recon_loss: 1348.7751; val_lambda_latent_contrastive_instanceloss: 6.3120; val_lambda_latent_contrastive_clusterloss: 2.0559; val_global_loss: 511997.6367; val_optim_loss: 511997.6367
Epoch 28/100 |█████---------------| 28.0% val_auroc_score: 0.9098; val_auprc_score: 0.9278; val_best_acc_score: 0.8501; val_best_f1_score: 0.8421; train_kl_reg_loss: 325.8071; train_edge_recon_loss: 4897.5165; train_gene_expr_recon_loss: 516324.5163; train_lambda_latent_adj_recon_loss: 1412.2301; train_lambda_latent_contrastive_instanceloss: 6.3396; train_lambda_latent_contrastive_clusterloss: 2.1076; train_global_loss: 522968.5126; train_optim_loss: 522968.5126; val_kl_reg_loss: 257.3006; val_edge_recon_loss: 4390.4759; val_gene_expr_recon_loss: 505502.5625; val_lambda_latent_adj_recon_loss: 1376.8004; val_lambda_latent_contrastive_instanceloss: 6.3225; val_lambda_latent_contrastive_clusterloss: 2.0599; val_global_loss: 511535.5234; val_optim_loss: 511535.5234
Epoch 29/100 |█████---------------| 29.0% val_auroc_score: 0.9089; val_auprc_score: 0.9271; val_best_acc_score: 0.8508; val_best_f1_score: 0.8424; train_kl_reg_loss: 327.5898; train_edge_recon_loss: 4894.8562; train_gene_expr_recon_loss: 515555.4314; train_lambda_latent_adj_recon_loss: 1412.6009; train_lambda_latent_contrastive_instanceloss: 6.3408; train_lambda_latent_contrastive_clusterloss: 2.1047; train_global_loss: 522198.9282; train_optim_loss: 522198.9282; val_kl_reg_loss: 257.2934; val_edge_recon_loss: 4386.9654; val_gene_expr_recon_loss: 503678.5039; val_lambda_latent_adj_recon_loss: 1391.2138; val_lambda_latent_contrastive_instanceloss: 6.3122; val_lambda_latent_contrastive_clusterloss: 2.0511; val_global_loss: 509722.3438; val_optim_loss: 509722.3438

Reducing learning rate: metric has not improved more than 0.0 in the last 4 epochs.
New learning rate is 1e-05.

Epoch 30/100 |██████--------------| 30.0% val_auroc_score: 0.9104; val_auprc_score: 0.9286; val_best_acc_score: 0.8522; val_best_f1_score: 0.8441; train_kl_reg_loss: 331.1566; train_edge_recon_loss: 4893.3296; train_gene_expr_recon_loss: 515422.0555; train_lambda_latent_adj_recon_loss: 1408.6650; train_lambda_latent_contrastive_instanceloss: 6.3405; train_lambda_latent_contrastive_clusterloss: 2.1040; train_global_loss: 522063.6516; train_optim_loss: 522063.6516; val_kl_reg_loss: 259.2392; val_edge_recon_loss: 4378.0033; val_gene_expr_recon_loss: 502574.3398; val_lambda_latent_adj_recon_loss: 1362.4006; val_lambda_latent_contrastive_instanceloss: 6.3134; val_lambda_latent_contrastive_clusterloss: 2.0721; val_global_loss: 508582.3711; val_optim_loss: 508582.3711
Epoch 31/100 |██████--------------| 31.0% val_auroc_score: 0.9077; val_auprc_score: 0.9256; val_best_acc_score: 0.8500; val_best_f1_score: 0.8414; train_kl_reg_loss: 329.9439; train_edge_recon_loss: 4893.6683; train_gene_expr_recon_loss: 515213.6679; train_lambda_latent_adj_recon_loss: 1410.4909; train_lambda_latent_contrastive_instanceloss: 6.3404; train_lambda_latent_contrastive_clusterloss: 2.1039; train_global_loss: 521856.2183; train_optim_loss: 521856.2183; val_kl_reg_loss: 261.1753; val_edge_recon_loss: 4394.0333; val_gene_expr_recon_loss: 500577.7383; val_lambda_latent_adj_recon_loss: 1404.9992; val_lambda_latent_contrastive_instanceloss: 6.3130; val_lambda_latent_contrastive_clusterloss: 2.0408; val_global_loss: 506646.2930; val_optim_loss: 506646.2930
Epoch 32/100 |██████--------------| 32.0% val_auroc_score: 0.9089; val_auprc_score: 0.9272; val_best_acc_score: 0.8512; val_best_f1_score: 0.8421; train_kl_reg_loss: 330.7787; train_edge_recon_loss: 4894.3714; train_gene_expr_recon_loss: 514896.1311; train_lambda_latent_adj_recon_loss: 1412.8600; train_lambda_latent_contrastive_instanceloss: 6.3401; train_lambda_latent_contrastive_clusterloss: 2.0960; train_global_loss: 521542.5765; train_optim_loss: 521542.5765; val_kl_reg_loss: 260.0209; val_edge_recon_loss: 4390.2610; val_gene_expr_recon_loss: 500822.7734; val_lambda_latent_adj_recon_loss: 1365.0766; val_lambda_latent_contrastive_instanceloss: 6.3141; val_lambda_latent_contrastive_clusterloss: 2.0634; val_global_loss: 506846.5078; val_optim_loss: 506846.5078
Epoch 33/100 |██████--------------| 33.0% val_auroc_score: 0.9100; val_auprc_score: 0.9279; val_best_acc_score: 0.8518; val_best_f1_score: 0.8438; train_kl_reg_loss: 332.0043; train_edge_recon_loss: 4892.1082; train_gene_expr_recon_loss: 514247.2500; train_lambda_latent_adj_recon_loss: 1409.5115; train_lambda_latent_contrastive_instanceloss: 6.3378; train_lambda_latent_contrastive_clusterloss: 2.1037; train_global_loss: 520889.3176; train_optim_loss: 520889.3176; val_kl_reg_loss: 261.5237; val_edge_recon_loss: 4383.5261; val_gene_expr_recon_loss: 502154.4727; val_lambda_latent_adj_recon_loss: 1393.9978; val_lambda_latent_contrastive_instanceloss: 6.3086; val_lambda_latent_contrastive_clusterloss: 2.0519; val_global_loss: 508201.8828; val_optim_loss: 508201.8828

Stopping early: metric has not improved more than 0.0 in the last 8 epochs.
If the early stopping criterion is too strong, please instantiate it with different parameters in the train method.
Model training finished after 13 min 27 sec.
Using best model state, which was in epoch 25.

--- MODEL EVALUATION ---
val AUROC score: 0.9083
val AUPRC score: 0.9251
val best accuracy score: 0.8500
val best F1 score: 0.8419
val MSE score: 2.5068

[10]:

# Compute latent neighbor graph
latent_key = 'garfield_latent'
sc.pp.neighbors(model.adata,
                use_rep=latent_key,
                key_added=latent_key)
# Compute UMAP embedding
sc.tl.umap(model.adata,
           neighbors_key=latent_key)

[11]:

# Compute latent Leiden clustering
latent_leiden_resolution = 1.2
latent_cluster_key = f"latent_leiden_{str(latent_leiden_resolution)}"
latent_key = "garfield_latent"

sc.tl.leiden(adata=model.adata,
             resolution=latent_leiden_resolution,
             key_added=latent_cluster_key,
             neighbors_key=latent_key)
len(model.adata.obs[latent_cluster_key].unique())

[11]:

21

Visualize Garfield Latent Space

[12]:

sc.settings.set_figure_params(dpi=100, facecolor='white')

sc.pl.umap(model.adata, color=['celltype_mapped_refined', latent_cluster_key], show=True, wspace=0.35, size=3)
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_16_0.png 
[15]:

## celltype
sc.pl.umap(model.adata, color=[ 'celltype_mapped_refined'], ncols=3,
           wspace=0.5, show=True, legend_loc='on data', size=1, legend_fontsize=3)
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_17_0.png 
[16]:

## batch information
sc.pl.umap(model.adata, color=['batch'], show=True)
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_18_0.png 
[17]:

## sample information
sc.pl.umap(model.adata, color=['sample'], ncols=3,
           wspace=0.5, show=True)
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_19_0.png 
[20]:

### Display spatial location distribution by sample
import squidpy as sq

batch_indeces = list(model.adata.obs['sample'].unique())
for batch_idx in batch_indeces:
    tmp = model.adata[model.adata.obs['sample'] == batch_idx, ]

    # Visualize cell-level annotated data in physical space
    sc.pl.embedding(tmp, basis="spatial",
                color=["celltype_mapped_refined"],
                ncols=1, wspace=0.20, edges=False)
    sc.pl.embedding(tmp, basis="spatial",
                color=[latent_cluster_key],
                ncols=1, wspace=0.20, edges=False)
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_20_0.png
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_20_1.png
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_20_2.png
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_20_3.png
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_20_4.png
../_images/tutorial_04.Garfield_spatia_molecular_niche_seqFISH_Mouse_Organogenesis_20_5.png 
[21]:

# Save trained model
model_folder_path = f"{workdir}/model"
os.makedirs(model_folder_path, exist_ok=True)

model.save(dir_path=model_folder_path,
           overwrite=True,
           save_adata=True,
           adata_file_name="adata_ref.h5ad")

Model saved successfully using pickle at /home/zhouweige/zhouwg_data/project/Garfield_tutorials/result/garfield_spRNA_embryo/model/attr.pkl

[ ]: