human-protein-atlas

Segmentation tells you where the cells are. The next layer up turns those pixels into meaning — a phenotypic fingerprint you can compare, cluster, and query. This season a proteome-aware vision model learned to read protein localization and morphology straight from Human Protein Atlas images, generalizing to new datasets it had never seen — while the field stayed honest about the wall it keeps hitting: models trained on one imaging setup break when the channels don’t match.

Cell morphology, learned from image-based spatial proteomics, is shown to predict molecular phenotypes — bridging single-cell imaging and proteomics.

ProtiCelli, a deep generative (diffusion) model, simulates microscopy images for 12,800 human proteins from just three cellular landmark stains — trained on the Human Protein Atlas — and powers Proteome2Cell, a step toward spatial virtual-cell modeling.

Large-scale immunofluorescence profiling maps how SARS-CoV-2 reshuffles the host protein landscape across 602 proteins, then screens approved drugs against the changed targets — yielding a 100,000-image public resource and pinpointing elesclomol as an antiviral modulator.