[QST] Why does the draw_graph result look so weird?

[Ci-TJ]

Thank you very much for developing rapids‑singlecell (rsc). While using rsc, I noticed that the results of its ForceAtlas2 (FA2) implementation differ significantly from those produced by Scanpy and cuGraph. Since I am not very familiar with the underlying principles of FA2, I would like to understand the reasons behind these differences.

start = time.perf_counter()
rsc.get.anndata_to_GPU(adata)
elapsed_min = (time.perf_counter() - start) / 60
print(f"{elapsed_min:.3f} min")

start = time.perf_counter()

SCVI_LATENT_KEY = "X_scVI"

# use scVI latent space for UMAP generation
#rsc.pp.neighbors(adata, use_rep=SCVI_LATENT_KEY)
#rsc.tl.umap(adata, min_dist=0.5)

rsc.pp.neighbors(adata, n_neighbors=30,use_rep="X_scVI",key_added="scVI")
rsc.tl.umap(adata,neighbors_key="scVI",key_added="X_umap_scVI")
rsc.tl.louvain(adata, resolution=0.6,neighbors_key="scVI",key_added="louvain_scVI")
rsc.tl.leiden(adata, resolution=0.6,neighbors_key="scVI",key_added="leiden_scVI")

elapsed_min = (time.perf_counter() - start) / 60
print(f"{elapsed_min:.3f} min")

start = time.perf_counter()
#rsc.tl.draw_graph(adata)
rsc.tl.draw_graph(adata, max_iter=500)
sc.pl.draw_graph(
    adata,
    color="Tissue",
    legend_loc="best",
    legend_fontsize="xx-small"
)
elapsed_min = (time.perf_counter() - start) / 60
print(f"{elapsed_min:.3f} min")

I used cuGraph to implement a simplified version of the FA2 algorithm.

start = time.perf_counter()
###################################
# Step 1: Extract the neighbor graph
A = adata.obsp["connectivities"].tocoo()

# Step 2: Convert to a cuDF DataFrame
df = cudf.DataFrame({
    "src": cp.asarray(A.row),
    "dst": cp.asarray(A.col),
    "weight": cp.asarray(A.data)
})

# Step 3: Build a cuGraph Graph object
G = cugraph.Graph()
G.from_cudf_edgelist(df, source="src", destination="dst", edge_attr="weight")

# Step 4: Run cuGraph ForceAtlas2
pos = cugraph.force_atlas2(G, max_iter=500)

###################################
# Step 5: Handle isolated nodes (cuGraph does not return nodes with degree = 0)
# Create a full coordinate matrix (n_cells × 2)
coords = np.full((adata.n_obs, 2), np.nan, dtype=np.float32)

# Sort by vertex index
pos_sorted = pos.sort_values("vertex")

# Fill coordinates for nodes returned by cuGraph
coords[pos_sorted["vertex"].to_pandas().values] = (
    pos_sorted[["x", "y"]].to_pandas().values
)

# Store coordinates in AnnData
adata.obsm["X_draw_graph_fa"] = coords

###################################
# Step 6: Make Scanpy recognize this layout
adata.uns["draw_graph"] = {
    "params": {"layout": "fa"}
}

###################################
# Step 7: Plot the graph layout
sc.pl.draw_graph(
    adata,
    color="Tissue",
    legend_loc="best",
    legend_fontsize="xx-small"
)

elapsed_min = (time.perf_counter() - start) / 60
print(f"{elapsed_min:.3f} min")

Best,
Seager

[Intron7]

Hey @Ci-TJ
I just came back from PTO yesterday. I fixed the graph creation from drawgraph can you please check if #529 fixes your issue?

[Ci-TJ]

Thanks for your help. For now, I struggling with the my server env. If I resolve other work, I will test these.