Method Spatial-MGCN

method/Spatial-mgcn/Spatial-mgcn.py

@@ -0,0 +1,254 @@
+#!/usr/bin/env python
+
+# Author_and_contribution: Niklas Mueller-Boetticher; created script
+# Author_and_contribution: Kirti Biharie; Added Spatial-MGCN
+
+import argparse
+
+CALC_ARI = False
+
+parser = argparse.ArgumentParser(
+    description="""Spatial-MGCN (https://doi.org/10.1093/bib/bbad262)"""
+)
+
+parser.add_argument(
+    "-c", "--coordinates", help="Path to coordinates (as tsv).", required=True
+)
+parser.add_argument(
+    "-m", "--matrix", help="Path to (transformed) counts (as mtx).", required=False
+)
+parser.add_argument(
+    "-f", "--features", help="Path to features (as tsv).", required=True
+)
+parser.add_argument(
+    "-o", "--observations", help="Path to observations (as tsv).", required=True
+)
+parser.add_argument(
+    "-n",
+    "--neighbors",
+    help="Path to neighbor definitions. Square matrix (not necessarily symmetric) where each row contains the neighbors of this observation (as mtx).",
+    required=False,
+)
+parser.add_argument("-d", "--out_dir", help="Output directory.", required=True)
+parser.add_argument(
+    "--dim_red",
+    help="Reduced dimensionality representation (e.g. PCA).",
+    required=False,
+)
+parser.add_argument("--image", help="Path to H&E staining.", required=False)
+parser.add_argument(
+    "--n_clusters", help="Number of clusters to return.", required=True, type=int
+)
+parser.add_argument(
+    "--technology",
+    help="The technology of the dataset (Visium, ST, ...).",
+    required=True,
+)
+parser.add_argument(
+    "--seed", help="Seed to use for random operations.", required=True, type=int
+)
+parser.add_argument(
+    "--config",
+    help="Optional config file used to pass additional parameters.",
+    required=False,
+)
+
+# Uncomment to calculate ARI every epoch as in original implementation
+if CALC_ARI:
+    parser.add_argument(
+        "-g", "--groundtruth",
+        help="Groundtruth.",
+        required=False,
+    )
+
+args = parser.parse_args()
+
+from pathlib import Path
+
+out_dir = Path(args.out_dir)
+
+# Output files
+label_file = out_dir / "domains.tsv"
+embedding_file = out_dir / "embedding.tsv"
+
+n_clusters = args.n_clusters
+technology = args.technology
+seed = args.seed
+
+## Your code goes here
+import json
+
+with open(args.config, "r") as f:
+    config = json.load(f)
+
+import random
+import numpy as np
+import pandas as pd
+import torch
+import tempfile
+import os
+import sys
+import scanpy as sc
+import scipy as sp
+import torch.optim
+import sklearn.cluster
+import sklearn.metrics
+import tqdm
+
+def get_anndata(args):
+    import anndata as ad
+
+    from PIL import Image
+    import scipy.io
+
+    X = sp.io.mmread(args.matrix)
+    if sp.sparse.issparse(X):
+        X = X.tocsr()
+
+    observations = pd.read_table(args.observations, index_col=0)
+    features = pd.read_table(args.features, index_col=0)
+
+    # Filter
+    if "selected" in observations.columns:
+        X = X[observations["selected"].to_numpy().nonzero()[0], :]
+        observations = observations.loc[lambda df: df["selected"]]
+    if "selected" in features.columns:
+        X = X[:, features["selected"].to_numpy().nonzero()[0]]
+        features = features.loc[lambda df: df["selected"]]
+
+    coordinates = (
+        pd.read_table(args.coordinates, index_col=0)
+        .loc[observations.index, :]
+        .to_numpy()
+    )
+
+    adata = ad.AnnData( # Rename spatial_pixel to spatial for Spatial-mgcn
+        X=X, obs=observations, var=features, obsm={"spatial": coordinates}
+    )
+
+    if args.image is not None:
+        adata.uns["image"] = np.array(Image.open(args.image))
+    else:
+        adata.uns["image"] = None
+
+    return adata
+
+
+adata = get_anndata(args)
+
+if CALC_ARI:
+    labels = pd.read_table(args.groundtruth, index_col=0)
+    adata = adata[adata.obs_names.isin(labels.index)]
+    labels = labels.loc[adata.obs_names]
+
+# Set seed
+random.seed(seed)
+torch.manual_seed(seed)
+np.random.seed(seed)
+torch.cuda.manual_seed(seed)
+
+# Work in a temprary folder
+with tempfile.TemporaryDirectory() as tmpdir:
+    gitdir = f"{str(tmpdir)}/Spatial-MGCN"
+
+    # Clone the repository to the specific commit
+    os.system(
+    f"""
+    git clone https://github.com/cs-wangbo/Spatial-MGCN.git {gitdir}
+    cd {gitdir} 
+    git reset --hard cf4412d
+    """)
+
+    # Set working directory as Spatial-MGCN directory
+    sys.path.append(f"{gitdir}/Spatial-MGCN")
+    import utils
+    import models
+
+    # Normalize data: min_cells, calculate HVG and scale
+    sc.pp.filter_genes(adata, min_cells=100)
+    sc.pp.highly_variable_genes(adata, flavor="seurat_v3", n_top_genes=config["fdim"])
+    adata = adata[:, adata.var['highly_variable']].copy()
+    adata.X = adata.X / np.sum(adata.X, axis=1).reshape(-1, 1) * 10000
+    adata.X = sp.sparse.csr_matrix(adata.X)
+    sc.pp.scale(adata, zero_center=False, max_value=10)
+
+    # Calculate graphs
+    fadj = utils.features_construct_graph(adata.X, k=config["k"]) 
+    sadj, graph_nei, graph_neg = utils.spatial_construct_graph1(adata, radius=config["radius"])
+
+    adata.obsm["fadj"] = fadj
+    adata.obsm["sadj"] = sadj
+    adata.obsm["graph_nei"] = graph_nei.numpy()
+    adata.obsm["graph_neg"] = graph_neg.numpy()
+
+    features = torch.FloatTensor(adata.X.todense())
+
+    nfadj = utils.normalize_sparse_matrix(fadj + sp.eye(fadj.shape[0]))
+    nfadj = sp.sparse.csr_matrix(nfadj)
+    nfadj = utils.sparse_mx_to_torch_sparse_tensor(nfadj)
+
+    nsadj = utils.normalize_sparse_matrix(sadj + sp.eye(sadj.shape[0]))
+    nsadj = sp.sparse.csr_matrix(nsadj)
+    nsadj = utils.sparse_mx_to_torch_sparse_tensor(nsadj)
+
+    graph_nei = torch.LongTensor(adata.obsm['graph_nei'])
+    graph_neg = torch.LongTensor(adata.obsm['graph_neg'])
+
+    # Create model
+    cuda = torch.cuda.is_available()
+
+    if cuda:
+            features = features.cuda()
+            nsadj = nsadj.cuda()
+            nfadj = nfadj.cuda()
+            graph_nei = graph_nei.cuda()
+            graph_neg = graph_neg.cuda()
+
+    model = models.Spatial_MGCN(nfeat=config["fdim"],
+                             nhid1=config["nhid1"],
+                             nhid2=config["nhid2"],
+                             dropout=config["dropout"])
+
+    if cuda:
+        model.cuda()
+
+    optimizer = torch.optim.Adam(model.parameters(), lr=config["lr"], weight_decay=config["weight_decay"])
+
+    # Train model
+    epoch_max = 0
+    ari_max = 0
+    idx_max = []
+    mean_max = []
+    emb_max = []
+
+    for epoch in tqdm.tqdm(range(config["epochs"])):
+        model.train()
+        optimizer.zero_grad()
+        com1, com2, emb, pi, disp, mean = model(features, nsadj, nfadj)
+        zinb_loss = utils.ZINB(pi, theta=disp, ridge_lambda=0).loss(features, mean, mean=True)
+        reg_loss = utils.regularization_loss(emb, graph_nei, graph_neg)
+        con_loss = utils.consistency_loss(com1, com2)
+        total_loss = config["alpha"] * zinb_loss + config["beta"] * con_loss + config["gamma"] * reg_loss
+        emb = pd.DataFrame(emb.cpu().detach().numpy()).fillna(0).values
+        total_loss.backward()
+        optimizer.step()
+
+        kmeans = sklearn.cluster.KMeans(n_clusters=args.n_clusters, n_init=10).fit(emb)
+        idx = kmeans.labels_
+
+        if CALC_ARI:
+            ari_res = sklearn.metrics.adjusted_rand_score(labels.to_numpy()[:,0], idx)
+            if ari_res > ari_max:
+                ari_max = ari_res
+                idx_max = idx
+                emb_max = emb
+        else:
+            idx_max = idx
+            emb_max = emb
+
+    # Write output
+    emb_df = pd.DataFrame(emb_max, index=adata.obs_names)
+    label_df = pd.DataFrame(idx_max, index=adata.obs_names, columns=["label"])
+
+    emb_df.to_csv(embedding_file, sep="\t", index_label="")
+    label_df.to_csv(label_file, sep="\t", index_label="")

method/Spatial-mgcn/Spatial-mgcn.yml

@@ -0,0 +1,19 @@
+channels:
+  - conda-forge
+  - defaults
+dependencies:
+  - python=3.8.13
+  - numpy=1.22.0
+  - pandas=1.4.4
+  - scipy=1.8.1
+  - matplotlib=3.5.3
+  - pytorch=1.11.0
+  - pip
+  - pip:
+      - stlearn==0.4.12
+      - torch_geometric==2.1.0
+      - torch_sparse==0.6.15
+      - torch_scatter==2.0.9
+      - git+https://github.com/cs-wangbo/Spatial-MGCN@cf4412df6db6276e0c8a3f036ff0b1661adfdb35
+      - python-louvain==0.16
+      - scikit-misc==0.2.0

method/Spatial-mgcn/Spatial-mgcn_optargs.json

@@ -0,0 +1,7 @@
+{
+    "matrix": "counts",
+    "integrated_feature_selection": true,
+    "image": false,
+    "neighbors": false,
+    "config_file": true
+}

method/Spatial-mgcn/config/config_1.json

@@ -0,0 +1,14 @@
+{
+    "epochs": 200, 
+    "lr": 0.001, 
+    "weight_decay": 5e-4, 
+    "k": 14, 
+    "radius": 560, 
+    "nhid1": 128, 
+    "nhid2": 64, 
+    "dropout": 0, 
+    "alpha": 1, 
+    "beta": 10, 
+    "gamma": 0.1, 
+    "fdim": 3000
+}