Many thanks to all authors - Arber Qoku, Martin Rohbeck, Florin Walter, Ilia Kats, Florian Buettner - and everyone who supported this work. Funding from @erc.europa.eu , @denbi.bsky.social and @dfg.de.
@dkfz.bsky.social @embl.org @oliverstegle.bsky.social

📑 Preprint: doi.org/10.1101/2025...
13/13

You’ll find all the code and documentation here:
mofaflex.readthedocs.io/stable/

pip install mofaflex

… and you’re ready to go! 12/n

Curious to build your very first MOFA-FLEX model and test it on your own data?

Check out the tutorial thread by @mlolab.bsky.social to see how it’s done: bsky.app/profile/mlol... 11/n

MOFA-FLEX is a unified framework that advances factor analysis for the omics community. Built on modular probabilistic programming, it lets researchers quickly design and deploy custom latent variable models by assembling reusable modules, fostering rapid innovation and extensibility. 10/n

3️⃣ We applied MOFA-FLEX to breast cancer data combining Xenium (166K cells; ~300-gene spatial) and Chromium (30K cells; ~3K-gene non-spatial) datasets from adjacent FFPE slides, revealing latent factors capturing global cell-type variation and transcriptome-wide spatially resolved gene programs. 9/n

2️⃣ When applying MOFA-FLEX to a multi-omic CITE-seq dataset from murine spleen and lymph nodes (N=14,870 cells), collected in two independent batches on different days, it was able to disentangle both batch and cell type variation. 8/n

1️⃣ We applied MOFA-FLEX to an scRNA-seq dataset of PBMCs from lupus patients (N=13,576 cells) treated with or without IFN-β, showing it can identify IFN-β gene programs consistent with known biology and reveal new program-linked genes (e.g., TNFSF10). 7/n

We showcase MOFA-FLEX in diverse applications: 1️⃣ robust recovery of gene programs from noisy prior knowledge in scRNA-seq, 2️⃣ disentangling technical and biological variation in multi-omic CITE-seq, and 3️⃣ spatial modelling revealing disease-linked gene programs in breast cancer. 6/n

Additionally, MOFA-FLEX adds a domain knowledge module that links latent factors to gene programs. It seamlessly integrates resources like Hallmark, Reactome, or custom sets to guide model learning and enhance the interpretability of discovered biological factors. 5/n

We teamed with @mlolab.bsky.social to develop MOFA-FLEX, addressing these challenges. Built on probabilistic programming, it unifies factor analysis extensions (flexible priors, non-negativity, supervision, alternative likelihoods), enabling declarative model design without manual engineering. 4/n

However, the rapidly evolving technological and experimental landscape demands universal, user-friendly frameworks that can be tailored to specific needs, such as incorporating spatial structure, temporal dynamics, noisy single-cell data, and domain-specific knowledge. 3/n

Advances in technology enabled parallel analysis of multiple biological layers. With the Buettner lab and others, we developed MOFA and related latent factor models to uncover key variation sources in multi-omics data. Today, factor models are first-line tools for sc and multi-omics analysis. 2/n