Matilda

Matilda is a multi-task neural network for single-cell multimodal omics. One model, trained once over RNA, ADT, and ATAC, drives classification, dimension reduction, feature selection, and simulation from a single shared representation. It is available in both Python (matilda-sc) and R (matilda), with matching results from either.

pip install "git+https://github.com/PYangLab/Matilda.git"   # Python (once on PyPI: pip install matilda-sc)
# or, in R:  remotes::install_github("PYangLab/Matilda", subdir = "matilda-r")

Get started How it works

1 Multimodal integration

Python Python tutorial Train and run every task with the object API, matilda.train() and matilda.task(), from the matilda-sc package.

R R tutorial Run the same model from a SingleCellExperiment with the R object API, reproducing the Python results bit-for-bit.

Concepts How it works See how per-modality encoders, the shared latent space, and the multi-task heads fit together in one network.

Reference API reference Full details on the Python functions and the R object methods that expose the shared model.

Multimodal integration Per-modality encoders for RNA, ADT, and ATAC feed a variational autoencoder whose shared latent space integrates every modality into one embedding.

One model, many tasks Because the encoders and latent space are shared, a single trained network performs classification, dimension reduction, feature selection, and simulation, and the tasks reinforce one another.

One tool, two interfaces Call the same model from Python or R: the object API in matilda-sc and in the R matilda package give bit-identical results on the same hardware.