Lab Newsletter — July 8, 2026: The Virtual Cell's Building Blocks

If the virtual cell is the destination, this week’s items are three of its load-bearing layers: the genome, the proteome, and the 3D tissue it all builds into. Each is getting an AI that turns cheap, available data into deep biology.

🧬 The genome gets its foundation models

Two models bracket the problem. Evo 2 (Arc Institute + NVIDIA, in Nature) is a generative genome model trained on 9.3 trillion nucleotides from 128,000+ genomes across 100,000+ species — it predicts whether BRCA1 mutations are pathogenic with >90% accuracy and can design synthetic genomes and even functional bacteriophages, all fully open-source. AlphaGenome (Google DeepMind, in Nature) comes at it from the other side: it reads up to 1 million base pairs at single-base resolution to predict regulatory effects — splicing, accessibility, RNA levels, TF binding — and scores a variant in ~a second, finally illuminating the non-coding 98% of the genome. Why it matters for the lab: a real virtual cell needs a genome layer that is both writable (Evo 2) and interpretable (AlphaGenome). These are that layer arriving.

🔬 A spatial protein map — from an ordinary slide

Spatial proteomics is powerful but expensive and hard to scale. A Stanford model, HEX (“H&E to protein expression”), sidesteps that by predicting 40 protein biomarkers directly from a standard H&E pathology slide — the cheapest, most ubiquitous image in medicine — producing virtual spatial-proteomics maps. Trained on 819,000 image tiles across 382 tumors, it improved lung-cancer prognostic accuracy by 22% and immunotherapy-response prediction by 24–39% over conventional biomarkers across 2,298 patients. Why it matters for the lab: this is our ProtiCelli thesis exactly — generate the molecular layer from imaging you already have. Turning a routine slide into a protein atlas is the image-to-omics bridge the virtual cell is built on.

🧫 Machine learning moves into 3D organoids

A new Trends in Biotechnology review maps how ML, AI and mathematical modeling are reshaping organoid research. Image-based readouts stay foundational — snapshot microscopy for morphology, and live-cell imaging for rich time series — and the review highlights deep visual proteomics revealing an in-vivo-like phenotype in transplanted human colon organoids. Why it matters for the lab: organoids are where a virtual cell meets real 3D tissue, and time-lapse imaging of living models is precisely what our self-driving microscope and REEF are built to generate at scale.

Read the genome, infer the proteome, grow the tissue — and model all three. None of these is the virtual cell on its own, but together they’re the scaffolding it will stand on.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.