Lab Newsletter — June 30, 2026: RNA's AlphaFold Moment, and a Proteome in Pictures

After a week heavy on agents and automation, today is about molecules and images — RNA finally getting its structure tools, pathology models maturing in the clinic, and a new proteome-scale picture of the cell from our own bench.

🧬 RNA catches up to proteins

Protein folding had its AlphaFold moment years ago; RNA is having its now. NuFold (Purdue; Nature Communications) predicts 3D RNA structure from sequence — its lead calls it “the RNA equivalent of AlphaFold” — and it’s open-source with a Colab notebook; trRosettaRNA2 (Nature Machine Intelligence, 2026) pushes accuracy further by fusing end-to-end learning with secondary-structure priors. On the design side, generative mRNA models (Science) now compose de novo sequences with enhanced translational capacity and stability — relevant to vaccines, protein-replacement and in-vivo cell therapies. Why it matters for the lab: the protein-folding playbook is transferring to a harder, more dynamic molecule — and structure + design together turn RNA from “read-only” into something you can engineer.

🔬 Pathology foundation models grow up at the clinical edge

A 2026 review maps how computational pathology is maturing: foundation models trained on enormous slide corpora — Virchow (0.95 AUC across 16 cancer types from ~1.5M H&E slides), Prov-GigaPath (1.3B tiles), the cytology-focused CytoFM, and KRONOS for spatial proteomics — now support subtyping, biomarker detection and pan-cancer tasks, alongside virtual staining (synthesizing diagnostic stains) and multimodal “copilots” (PathChat, TeamPath). Real products are FDA-cleared and adoption is climbing (~10% of US labs, 2024). Why it matters for the lab: this is the regulated, clinical edge of the image × omics work we care about — virtual staining echoes our generative-imaging direction, and KRONOS sits right on the spatial-proteomics bridge.

📖 From the lab: a proteome-wide image of the cell

Hot off bioRxiv, a new paper with the lab’s Wei Ouyang (and Emma Lundberg) — ProtiCelli — uses a deep generative (diffusion) model to simulate microscopy images for 12,800 human proteins from just three landmark stains, trained on the Human Protein Atlas. It then generates Proteome2Cell: ~30.7M images forming 2,400 “virtual cells” across 12 cell lines, recovering subcellular organization, protein–protein interaction landscapes and even drug-induced changes from morphology alone. Why it matters for the lab: it’s a concrete step toward spatial virtual-cell modeling — turning spatial proteomics from cataloguing proteins into simulating whole cellular systems, exactly the horizon this newsletter keeps circling.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. Have lab news to share — a talk, paper, conference or release? Message me on Slack.