Handling Different Feature Dimensions Across Batches in feat_dims.toml File

[EdenLiyy]

Hello,

I am using the feat_dims.toml file, which contains the feature dimensions for each modality. However, I encountered an issue: since I applied filtering criteria to the same modality (e.g., ATAC), the feature dimensions for ATAC are different across batches.

I would like to know if the model can handle this scenario. Specifically, can the model work with the situation where feature dimensions for the same modality (e.g., ATAC) vary between batches, or does the model require all batches to have consistent feature dimensions?

Additionally, I noticed that the feature number for ATAC is based on a chromosome statistical feature count list. How is this list generated? Is it the sum of the feature counts across all batches?

Key points:
I am using a single feat_dims.toml file to manage feature dimensions for all modalities.

For ATAC modality, I applied filtering criteria that caused feature dimensions to vary across batches.

Can the model handle this configuration, where the same modality (e.g., ATAC) has different feature dimensions across different batches?

For the ATAC modality, is the feature count list based on chromosome statistics computed as the sum of feature counts across all batches?

[JiahaoChow]

Hi,

Thanks for your interest!

✅ Handling Varying Feature Dimensions Across Batches

We do support scenarios where different batches have different feature availability — including for the ATAC modality. This is managed via mask files, where:

1 indicates a present feature
0 indicates a missing feature

This mechanism is already used in our handling of RNA and ADT. Even though our paper assumes consistent ATAC features across batches (refer to preprocessing), the codebase supports masked input for ATAC as well.

📄 About feat_dims.toml and ATAC Feature Splitting

In your feat_dims.toml, a line like:

atac = [100, 200, 300, ...]

means the ATAC matrix will be split into segments with 100, 200, 300 features, respectively — typically aligned with chromosomes. To work correctly, make sure:

• Your feature matrix is sorted by chromosome (e.g., chr1_xx, chr1_yy, chr2_xx, ...)
• This ordering is consistent across all batches
• All matrices are padded to the same full dimension, with masks indicating missing entries

✅ Summary

Yes, the model can handle feature dimension variations across batches via masking.
Just ensure all matrices are aligned to the same shape and ordered consistently.
The feat_dims.toml must reflect the global dimension breakdown, not per-batch shapes.

[EdenLiyy]

Thank you for your response. Regarding the integration of three-modal data, there is currently no missing modality, but integration is needed. About the mentioned use of masks to indicate missing entries, does it specifically refer to the content of the corresponding modality in the mask folder? Could you please elaborate on the specific content of the files I need to provide?

[JiahaoChow]

For each batch and modality, an optional mask file can be provided in .csv format under the batch_x/mask/ folder. Please refer to our documentation for more details.