Phenom-Beta is a vision-transformer model that learns biologically meaningful features from cellular microscopy images. Its training dataset contains 2.2 million images spanning thousands of compound or genetic perturbations across multiple cell types. section 13.2.1
The model divides each image into square patches, embeds the patches as tokens, and processes the token sequence through transformer layers to build an image representation. section 13.2.1
In phenotypic screening, compounds are evaluated by their effects on cell shape and behavior. A challenge is deconvolution: identifying the cellular targets responsible for those effects. Phenom’s learned features can connect observed phenotypes with possible molecular mechanisms. section 13.2.1
A possible kinase inhibitor
If a new compound produces a cellular phenotype similar to that of a known kinase inhibitor, Phenom’s features might suggest that the new compound also inhibits kinases. The similarity is a clue, not an established mechanism. section 13.2.1
