Lab Newsletter — July 13, 2026: Wiring Diagrams for the Cell

A cell isn’t a parts list; it’s a wiring diagram. Today’s items are three attempts to draw it — the physical connections, the regulatory logic, and the hard question of whether the drawings are true.

🕸️ The interactome goes public, at scale

The AlphaFold Database just took a big step from single proteins to complexes. In a collaboration between EMBL-EBI, Google DeepMind, NVIDIA and Seoul National University, some 30 million predicted complexes were computed — 1.7 million high-confidence homodimers (plus 18M lower-confidence for bulk download), and, as of a 19 May update, nearly 80,000 high-confidence heterodimers with 8.1 million more available. Free to everyone, it would have cost ~17 million GPU-hours to reproduce. As the team frames it, this is “a first step toward a full description of the human interactome” — because 20,000 proteins produce their staggering complexity mostly through how they interact. Why it matters for the lab: interactions are where biology hides, and recovering the protein-interaction landscape is exactly what our ProtiCelli work does from images — now there’s an open structural map to triangulate against.

🧬 Interpretable models for the regulatory wiring

Structure is one layer; regulation is the other. New single-cell foundation models are trying to map it without becoming black boxes. CellVQ (Nature Communications) reports beating scGPT and scFoundation on perturbation and annotation tasks while adding an interpretable graph view (CellVQ-Graph) for gene-regulatory-network analysis — reading out which genes drive a cell state, not just predicting it. In a similar spirit, Novartis’s CellxPert critiques the common trick of simulating a knockout by shuffling gene-expression tokens (which shoves models out of distribution) and instead builds molecular→cellular→multicellular layers to keep perturbations biologically grounded. Why it matters for the lab: an interpretable regulatory model is the difference between a virtual cell you can trust and one you can only admire.

⚖️ Are the wiring diagrams real?

The sobering counterpoint keeps the field honest. A 2026 evaluation finds that the attention in these single-cell models often captures co-expression rather than unique regulatory signal — correlation dressed as causation — and perturbation predictors still struggle to clearly beat simple linear baselines. That’s why benchmarks like PertEval-scFM exist. Why it matters for the lab: a wiring diagram is only useful if its edges are real, and the way you find out is the same as always — perturb, observe, validate. It’s the loop REEF is built to close and the discipline our virtual-cell work has to hold.

Map the connections, map the logic, then check the map against the cell. The interactome is finally becoming a public object — and the real work is making sure the lines we draw across it are true.

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.