Lab Newsletter — September 26, 2026: Mapping the Cell's Wiring

Yesterday we watched a microscope read a cell’s future from its images; earlier this week we named cells and folded proteins. Today we connect them: a cell isn’t a bag of independent parts, it’s a network — proteins that touch, bind, and assemble into molecular machines. Knowing which proteins interact, and what they build together, is the cell’s wiring diagram — and a load-bearing layer of the virtual cell. Today’s digest is about the AI now drawing that diagram at proteome scale.

🗺️ The reference map

You can’t study a network without a map of it. STRING (Szklarczyk et al., Nucleic Acids Research, 2019) is the field’s, built on the premise that “proteins and their functional interactions form the backbone of the cellular machinery” whose “connectivity network needs to be considered for the full understanding of biological phenomena.” STRING sets out “to collect, score and integrate all publicly available sources of protein-protein interaction information” and “to achieve a comprehensive and objective global network, including direct (physical) as well as indirect (functional) interactions,” now spanning “5090” organisms. An open, scored atlas of the interactome — the same open-infrastructure spirit as the lab’s BioImage Model Zoo, applied to interactions.

🧬 Read interaction from evolution

Where does a new interaction signal come from? Evolution leaves one. Cong et al. (Science, 2019) mined “coevolution between 5.4 million pairs of proteins in Escherichia coli” (and 3.9 million in M. tuberculosis): proteins that must fit together tend to mutate in concert, and that coupling, plus structure modeling, “predict[s] protein-protein interactions (PPIs) with an accuracy that benchmark studies suggest is considerably higher than that of proteome-wide two-hybrid and mass spectrometry screens.” The result was discovery, not just recapitulation — “hundreds of previously uncharacterized PPIs” that “add components to known protein complexes … and establish the existence of new ones.”

🔤 Predict from sequence alone

Coevolution needs deep alignments; sequence-based deep learning can go further and faster. Sledzieski et al. (Cell Systems, 2021) built D-SCRIPT, “an interpretable and generalizable deep-learning model, which predicts interaction between two proteins using only their sequence and maintains high accuracy with limited training data and across species.” Impressively, “the inter-protein contact map output by D-SCRIPT has significant overlap with the ground truth” — it learns where proteins touch, not just whether — letting it “screen for PPIs” genome-wide in species like cow where almost no interaction data exist. Structure-aware, but structure-free at inference: the same representation- learning wager the lab makes across biology.

🕸️ Add the network view

A protein’s interactions aren’t independent — the shape of the whole network is itself a clue. Singh et al. (Bioinformatics, 2022) unified the two schools with Topsy-Turvy, synthesizing “bottom-up” sequence features and “top-down” network patterns in one model. It delivers “genome-scale, interpretable PPI prediction for non-model organisms with no existing experimental PPI data,” and — crucially for anyone running these at scale — “running Topsy-Turvy … screens is feasible for whole genomes, and thus these methods scale to settings where other methods (e.g. AlphaFold-Multimer) might be infeasible.” Accuracy you can actually afford across a proteome.

🧩 Fold the complex directly

When you can afford structure, it pays off. Bryant, Pozzati & Elofsson (Nature Communications, 2022) — at Stockholm University / SciLifeLab, the lab’s own backyard — turned AlphaFold2 onto interactions, applying it “for the prediction of heterodimeric protein complexes.” With “optimised multiple sequence alignments,” it produced “models with acceptable quality (DockQ ≥ 0.23) for 63% of the dimers,” and, cleverly, they built “a simple function to predict the DockQ score” that distinguishes “interacting from non-interacting proteins with state-of-art accuracy” — recovering “51% of all interacting pairs at 1% FPR.” Structure prediction becomes an interaction detector.

🏗️ The proteome-scale payoff

Put coevolution and deep folding together and you can rebuild a cell’s machines wholesale. Humphreys et al. (Science, 2021) combined “proteome-wide amino acid coevolution analysis and deep-learning–based structure modeling” (RoseTTAFold + AlphaFold) to screen “8.3 million pairs of yeast proteins, identify 1505 likely to interact, and build structure models for 106 previously unidentified assemblies and 806 that have not been structurally characterized.” These complexes, “as many as five subunits,” touch “almost all key processes in eukaryotic cells” — a first structural draft of a eukaryote’s molecular machinery.

🧫 Why it’s our kind of problem

Read across the six and it’s the wiring layer of the virtual cell: you cannot simulate a cell without knowing which proteins interact and what they build. Two lab themes recur. First, the structure-based vs. sequence/network-based trade-off — AlphaFold-quality complexes are accurate but heavy, while D-SCRIPT and Topsy-Turvy scale to whole genomes and orphan species — is precisely the case for the lab’s BioEngine: make the heavy methods runnable at scale, for everyone. Second, open shared maps (STRING) and learned representations are the same bets the lab makes across structure, imaging, and single cells. And it’s close to home: the AlphaFold-for-interactions work comes from SciLifeLab, down the road. A parts list was never enough — biology runs on the connections, and we’re finally able to draw them.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.