VirTues-M2
[!NOTE] As shown in the GitHub repository, the latest model and it's downstream task heads are currently in the 'v2' branch.
VirTues-M2 is a foundation model for tissue biology that jointly learns from routine H&E histopathology and multiplex spatial proteomics within a single architecture. It follows an early-fusion design: pixel-aligned images are embedded by modality-specific tokenizers (an RGB tokenizer for H&E and a multiplex tokenizer for spatial proteomics) and processed by a shared vision transformer backbone. In the multiplex tokenizer, each channel is combined with a protein foundation model embedding of its marker identity. This lets VirTues-M2 handle arbitrary and varying marker panels and stay robust to missing modalities, incomplete pairing and imperfect alignment. VirTues-M2 is pretrained on the largest open-source multiplex spatial proteomics collection with partially paired H&E, covering 5,849 patients, 37 studies and 17,275 multiplex tissues acquired with PhenoCycler (CODEX), Orion, IMC and MIBI. The resulting joint representation can be used in tasks such as treatment response, tissue phenotyping, panoptic cell segmentation, virtual staining, biomarker discovery and adaptive marker panel selection.
For model usage and installation instructions, see the [GitHub] repository.