One AI model learns to generate fake tissue slides from molecular data — and does it better than specialist tools

Researchers built a single AI model that learns across three types of medical data at once: tissue images, gene expression, and clinical notes. Instead of needing separate tools to translate between each pair of data types, this model handles all of them together, which means it can fill in missing pieces (like a missing stain or gene test) more accurately than specialized models designed for one specific task.

Pathology diagnosis often requires multiple tests on limited tissue samples. A single test is expensive and tissue is scarce, so doctors frequently have only partial data. Until now, hospitals had to use task-specific tools to guess what the missing data looked like — and those guesses were mediocre. This model reduces the guessing error by 23 to 50 percent depending on the task. That matters because bad guesses in pathology can miss disease signals or create false ones. The real question is whether hospitals will trust a hallucinated stain or gene pattern enough to use it in actual diagnosis, or whether it stays a research tool.