Multimodal perturbation readouts (Perturb-CITE-seq, Perturb-ATAC, multiome)

Context

pertpy is largely RNA-centric, even though several bundled workflows use multimodal data (e.g. the Papalexi Perturb-CITE-seq dataset used as an example for Mixscape and perturbation space).
CRISPR screens with protein (Perturb-CITE-seq), chromatin (Perturb-ATAC), and joint multiome readouts are increasingly common.

What's missing

• Tools that operate on protein / ATAC (or joint) modalities rather than assuming an RNA .X — e.g. distances, perturbation efficacy, and perturbation-space representations over non-RNA or multimodal features.
• Guide/perturbation assignment and QC that leverage all modalities.
• Consistent MuData-first handling across tools.

Proposal / API

Make the modality explicit where tools currently assume RNA (layer / obsm / modality argument), and validate on Perturb-CITE-seq (protein) and a Perturb-ATAC / multiome dataset.
Reuse the MuData conventions already present (Mixscape / Milo load into MuData).

Start by reviewing the existing Mixscape and Milo workflows and the MuData conventions they use, then identify where tools assume an RNA .X. Validate the modality-explicit behavior on the Papalexi Perturb-CITE-seq dataset and a Perturb-ATAC or multiome dataset; done means protein, ATAC, and joint modalities are supported consistently where applicable.