Journal: bioRxiv

Article Title: Generative machine learning unlocks the first proteome-wide image of human cells

doi: 10.64898/2026.03.31.715748

Figure Lengend Snippet: ( A ) Left : Protein–protein interaction network centered on P4HA2, illustrating its context-dependent interactions with ER- and nucleoplasm-localized proteins. Right : Real P4HA2 images (red border) and ProtiCelli -generated images of ten interaction partners for two representative cells, grouped by interaction context (gold: ER/collagen; blue: mRNA processing). The rightmost column shows the binarized spatial interaction overlap between P4HA2 and the mean projection of proteins within partner groups. Scale bar, 10 µm. ( B ) Representative real and ProtiCelli -generated fluorescence microscopy images for four proteins (MECP2, BABAM2, YWHAB, BPTF) under Paclitaxel, untreated, and Vorinostat conditions, demonstrating the model’s ability to capture single-cell expression variance across treatment contexts. Scale bar, 10 µm. ( C ) Boxplots comparing ProtiCelli -generated versus real normalized protein intensities for ten representative proteins across treatment conditions. Median values are annotated; statistical significance (Mann-Whitney U test) is indicated by bracket annotations. ( D ) Scatter plots comparing average foreground intensity between real and ProtiCelli -generated images for twelve cell cycle-regulated proteins. Each point represents one single cell and the red line denotes the linear fit with 95% confidence interval. Pearson correlation coefficients ( r ) are reported per protein. ( E ) Representative real and ProtiCelli -generated immunofluorescence images of the FUCCI markers CDT1 (A-549) and GMNN (MDA-MB-467). Protein signal is shown in green; microtubule staining (red), derived from real images, is identical across all panels. Scale bar, 10 µm. ( F, G ) Generated versus measured FUCCI marker intensities of CDT1 and Geminin (GMNN) across cell cycle stages (G1, G1/S, G2) without fine-tuning, after fine-tuning, and for real images, in A-549 ( F ) and MDA-MB-468 ( G ) cell lines. Fine-tuned ProtiCelli accurately capture cell-cycle-dependent expression dynamics for both markers. Significance was determined using the Mann-Whitney U test.

Article Snippet: Large-scale initiatives such as the Human Protein Atlas (HPA) have systematically mapped protein expression and subcellular localization across human cell types using immunofluorescence (IF) microscopy, imaging one protein at a time alongside reference markers that delineate cellular landmarks ( , ).

Techniques: Generated, Fluorescence, Microscopy, Single Cell, Expressing, MANN-WHITNEY, Immunofluorescence, Staining, Derivative Assay, Marker