Current multimodal single-cell integration methods act as black boxes, lacking meaningful interpretability. We introduce NetworkVI, a VAE that performs integration via gene-gene interactions and the Gene Ontology for biologically grounded analysis.

RT @razoralign: NetworkVI: Biologically Guided Variational Inference for Interpretable Multimodal Single-Cell Integration and Mechanistic D….

Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.

NetworkVI is a group effort by Julius Upmeier zu Belzen, Luis Herrmann, Khue Nguyen, supervised by @fabian_theis & @nebww & @CaptainSysBio at @ChariteBerlin and @HelmholtzMunich. Code: .Preprint: .🧠 Open-source & ready to use!.

We demonstrate how NetworkVI identifies immune evasion mechanisms via GO programs in a CRISPR-perturbed melanoma dataset not detectable with standard methods. This highlights the medical utility of interpretable multimodal models.

NetworkVI also incorporates a GO-specific covariate attention mechanism, allowing the model to adjust for sample-level metadata (e.g., donor or condition), improving both performance and interpretability in heterogeneous datasets.

These scores reveal associations between genes and GO terms, helping to identify active biological processes and their regulatory drivers in a given condition or perturbation. This opens a path toward mechanistic interpretation beyond enrichment tests.

NetworkVI’s architecture incorporates prior biological knowledge:. 1. Gene-gene interactions from TADs .2. The structure of the Gene Ontology .This enables inference of gene and GO term importance scores at modality- and cell-specific resolution.

📊 NetworkVI achieves state-of-the-art performance on bimodal and trimodal data: .✅ Accurate integration .✅ Imputation of missing modalities .✅ Query-to-reference mapping .✅ Interpretation of cell type– and perturbation-specific signatures.

RT @Tillmann_Rhe: Very happy to see this online after the presentation at the @imimicworkshop last year! .In this paper, we propose Seg-Hi….

RT @NobelPrize: BREAKING NEWS.The Royal Swedish Academy of Sciences has decided to award the 2024 #NobelPrize in Chemistry with one half to….

RT @sid_thesci_kid: new paper drop in @NatureBiotech .It was a journey and half getting this out, and I couldn’t have done it w/o the kinde….

This project was co-led by @sid_thesci_kid, @JacobMGershon, and Jeremiah Sims, from whom I learned a lot and am very grateful for the work we did together at @UWproteindesign.

I am particularly excited about our experiments exploring the fitness landscape of the protein GB1 via an iterative gradient-based guidance process on very sparse experimental data. Thus, ProteinGenerator may be useful for machine-learning-assisted directed evolution campaigns.

We demonstrate how joint modeling of sequence and structure, can be employed for scaffold design, multistate design, and activity guidance, among other things. Please also check out Sam’s thread, which shows so much more cool stuff.

Thrilled to share our @NatureBiotech paper “Multistate and functional protein design using RoseTTAFold sequence space diffusion”

RT @SamTipps: 🥁 Our manuscript "Multistate and functional protein design using RoseTTAFold sequence space diffusion" is finally out @Nature….

RT @NatureBiotech: Multistate and functional protein design using RoseTTAFold sequence space diffusion https://t.co….

RT @fengel97: Last day of #ERCL23 🧬

RT @Jas_Hughes: The fact that there are multiple candidates for what this linear regression is blows my mind 🤯. (Original, of course: https….

Also want to briefly emphasize my favorite finding: The model can be guided via for instance amino acid weights, motifs or classifier gradients. However, introducing GMMs instead of a Normal as the noise sampeling prior in diffusion cause the generation of more diverse proteins.