WNN step error: 'Fewer than 20 cells with Jaccard sim > 0' in COSMOS

[Zhihong-Lin]

Thank you for developing and open-sourcing COSMOS. It has been very helpful for our spatial multi-omics analysis. But when training reaches the WNN stage (w == 1), the program exits with the following error:

Fewer than 20 cells with Jacard sim > 0

Dataset summary

• Modalities: RNA + CODEX
• Number of cells: ~264,000
• RNA features: ~6,000 genes
• CODEX features: 16 protein markers
• Spatial coordinates available for all cells

Workflow

cosmos_comb = cosmos.Cosmos(adata1=adata1, adata2=adata2)
cosmos_comb.preprocessing_data(n_neighbors=10)

cosmos_comb.train(
    spatial_regularization_strength=0.05,
    z_dim=10,
    lr=1e-3,
    wnn_epoch=500,
    total_epoch=1000,
    random_seed=seed,
    gpu=0
)

I tested different settings (e.g. z_dim from 50 to 10, n_neighbors from 10 to 60 ), but the error persists whenever WNN is executed.

I am wondering whether this behavior could be related to the large difference in feature dimensionality between the two modalities (RNA with ~6,000 genes versus CODEX with 16 protein markers). In large-scale datasets, could this imbalance lead to very sparse KNN overlap during the WNN step, and thus trigger the Jaccard-based bandwidth estimation error?

Thank you very much for your time and for making COSMOS publicly available. I look forward to your insights.