API Reference

Core

`sparank.SpaRank`	High-level pipeline: register modalities -> prepare -> fit -> predict.
`sparank.ExpConfig`	Unified configuration covering both unimodal and multimodal setups.
`sparank.ModalityConfig`	Specification for one omics modality (RNA, ADT, ATAC, ...).
`sparank.SimulationConfig`	Parameters for pseudo-spot simulation via SPACEL.

`sparank.modules.SpotRankTransformer`	Unified 1-to-N modality Transformer for spatial deconvolution.
`sparank.modules.FusionLayer`	Fuse N modality embeddings into a single vector.
`sparank.modules.NTXentLoss`	Normalised Temperature-scaled Cross-Entropy (NT-Xent) loss for contrastive learning on paired views.
`sparank.modules.DeconvCrossEntropy`	Cross-entropy loss for cell-type proportion deconvolution.

`sparank.data.build_vocab`	Build per-modality vocabularies.
`sparank.data.tokenize_batch`	Tokenise an AnnData batch across 1-to-N modalities.
`sparank.data.MemmapDataset`	Memory-mapped dataset for 1-to-N modalities.
`sparank.data.InferenceDataset`	Minimal in-memory dataset designed for inference.
`sparank.data.simulate`	Batch-proportional pseudo-spot simulation with memmap writing.
`sparank.data.find_sc_markers`	Batch-aware marker gene detection using scanpy's rank_genes_groups.
`sparank.data.normalize_rna`	Library-size normalisation followed by log1p, stored as a layer.

Train a SpotRankTransformer with CLS + CL + MRP.