🧬 𝐖𝐡𝐚𝐭 𝐢𝐟 𝐀𝐈 𝐜𝐨𝐮𝐥𝐝 𝐫𝐞𝐚𝐝 𝐚 𝐬𝐥𝐢𝐝𝐞 𝐨𝐟 𝐜𝐚𝐧𝐜𝐞𝐫 𝐭𝐢𝐬𝐬𝐮𝐞 𝐚𝐧𝐝 𝐢𝐧𝐬𝐭𝐚𝐧𝐭𝐥𝐲 𝐦𝐚𝐩 𝐢𝐭𝐬 𝐚𝐜𝐭𝐢𝐯𝐞 𝐠𝐞𝐧𝐞 𝐞𝐱𝐩𝐫𝐞𝐬𝐬𝐢𝐨𝐧?
At Portrai, our mission is to decode complex diseases with spatial intelligence. Today, we’re taking a major step forward: introducing VGL-stGen—a multimodal generative foundation model that connects tissue morphology, gene expression, and cell identity into one shared framework.
🔬 𝐎𝐧𝐞 𝐌𝐨𝐝𝐞𝐥. 𝟏𝟑 𝐕𝐞𝐫𝐬𝐚𝐭𝐢𝐥𝐞 𝐒𝐩𝐚𝐭𝐢𝐚𝐥 𝐁𝐢𝐨𝐥𝐨𝐠𝐲 𝐓𝐚𝐬𝐤𝐬.
Instead of building isolated models for specific tasks, a single VGL-stGen checkpoint supports 13 distinct downstream tasks—from H&E-based tissue and cell-type classification to gene-list generation and expression completion.
Key Highlights:
• 𝐎𝐮𝐭𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐞𝐝 𝐋𝐚𝐫𝐠𝐞𝐫 𝐌𝐨𝐝𝐞𝐥𝐬: Surpassed larger off-the-shelf multimodal LLMs across key benchmarks under matched evaluation protocols.
• 𝐏𝐨𝐰𝐞𝐫𝐞𝐝 𝐛𝐲 𝐏𝐨𝐫𝐭𝐫𝐚𝐢𝐀𝐓𝐋𝐀𝐒: Trained on a dataset combining 69 paired ST+H&E samples with single-cell references and 30k+ cancer patches.
• 𝐃𝐞𝐞𝐩 𝐁𝐢𝐨𝐥𝐨𝐠𝐲 𝐯𝐢𝐚 𝐒𝐞𝐥𝐟-𝐒𝐮𝐩𝐞𝐫𝐯𝐢𝐬𝐢𝐨𝐧: Bridges visual histology directly with molecular cellular states.
This proof-of-concept study in lung cancer marks a crucial foundation toward our longer-term vision: in silico virtual human test-driven drug development.
Explore the interactive showcase, paper, and model weights below:
https://lnkd.in/gEShfrm4
Read the preprint:
https://lnkd.in/gSm_hjZM
Model:
https://lnkd.in/gDbpJtUF
