M3

A deep generative framework that learns factorised, condition-aware embeddings for multimodal, multi-condition, multi-sample single-cell data, separating biological signal from condition and batch effects so integration, patient-level inference, and attribution stay internally consistent.

pip install m3-sc

Get started How it works

1 Factorised dimension reduction

Integrate Representation learning Learn factorised, condition-aware cell embeddings that integrate across batches, conditions, and modalities while preserving condition signal.

Predict Patient-level inference Aggregate cell embeddings into pseudo-bulk profiles to predict a sample's condition, such as disease status, on unseen query data.

Interpret Feature attribution Attribute a condition back to genes, individual cells, and cell types for multi-resolution interpretation.

Generate Data augmentation Synthesise new donors and cells to augment small cohorts directly from the learned representation, with batch-stratified control.

Factorised, condition-aware M3 learns embeddings that separate three sources of variation explicitly, biological signal, condition effect, and batch effect, so batch correction does not overcorrect the signal you actually care about.

Multimodal and mosaic A product-of-experts fuses modality-specific encoders into one shared biological embedding per cell, handling both fully observed and mosaic designs where some samples are missing a modality.

Interpretable across tasks Every result decodes from one shared, condition-aware representation, keeping integration, patient-level inference, generation, and gene-, cell-, and cell-type-level attribution internally consistent.