Lab Newsletter — September 11, 2026: Medicine as a Graph

For a week this digest has zoomed in — on a single molecule at a time. We designed proteins from scratch, evolved the ones we have, computed the forces that hold them together, and screened chemical space for antibiotics. Today we zoom all the way out. Because biology isn’t only a pile of molecules — it’s a web of relationships: this gene regulates that pathway, this drug hits that target, this disease shares mechanisms with that one. Draw all of it as one enormous graph, and a new question opens up: which edges are missing? Which drug already on a pharmacy shelf might treat a disease no one has tried it on? Today’s digest is about teaching machines to reason over that graph — and the hidden medicine they find in it.

🕸️ The founding move: integrate everything, then walk it

The idea starts with a refusal to look at one thing in isolation. Himmelstein et al. (eLife, 2017) set out to make drug repurposing computable: “the ability to computationally predict whether a compound treats a disease would improve the economy and success rate of drug approval.” Their move was to fuse the scattered literature into a single network. They “constructed Hetionet … an integrative network encoding knowledge from millions of biomedical studies” — a graph that in v1.0 held “47,031 nodes of 11 types and 2,250,197 relationships of 24 types,” wiring together compounds, diseases, genes, pathways, side effects and more. Then they learned which patterns of connection distinguish a real treatment from a coincidence, and used them to score “the probability of treatment for 209,168 compound-disease pairs.” The detail that makes it our kind of science: the whole project “was entirely open and received realtime feedback from 40 community members.” Medicine, redrawn as a graph anyone could read.

🔗 Graph neural networks enter: predict the exact missing edge

Hetionet walked the graph with hand-designed path patterns; the next step was to learn the patterns. Zitnik, Agrawal & Leskovec (Bioinformatics, 2018) brought deep learning to the network with Decagon, aimed at a real clinical danger — the side effects that emerge when drugs are combined. They “construct[ed] a multimodal graph of protein-protein interactions, drug-protein target interactions and the polypharmacy side effects, which are represented as drug-drug interactions, where each side effect is an edge of a different type,” then built “a new graph convolutional neural network for multirelational link prediction in multimodal networks.” The payoff over prior methods was specificity: Decagon “can predict the exact side effect, if any, through which a given drug combination manifests clinically,” and it does so accurately, “outperforming baselines by up to 69%.” Not just whether two drugs clash, but how — a missing edge, named.

🦠 The method meets an emergency

A framework proves itself under pressure. When COVID-19 arrived and there was no time for de novo discovery, Morselli Gysi et al. (PNAS, 2021) turned the network loose on the problem of repurposing. They “deployed algorithms relying on artificial intelligence, network diffusion, and network proximity, tasking each of them to rank 6,340 drugs for their expected efficacy against SARS-CoV-2.” A key lesson was humility about any single model — “a consensus among the different predictive methods consistently exceeds the performance of the best individual pipelines.” And the graph earned its keep at the bench: screening the top-ranked drugs in human cells gave “a 62% success rate, in contrast to the 0.8% hit rate of nonguided screenings.” The most striking finding is why it worked — “76 of the 77 drugs that successfully reduced viral infection do not bind the proteins targeted by SARS-CoV-2, indicating that these network drugs rely on network-based mechanisms that cannot be identified using docking-based strategies.” Exactly the medicine that yesterday’s structure-based screens can’t see.

🗺️ An open map for precision medicine

A model is only as good as the graph beneath it — so the next contribution was a better graph, released for everyone. Chandak, Huang & Zitnik (Scientific Data, 2023) built PrimeKG, “a multimodal knowledge graph for precision medicine analyses.” It “integrates 20 high-quality resources to describe 17,080 diseases with 4,050,249 relationships representing ten major biological scales” — from protein perturbations and pathways up to anatomy and clinical phenotype. Crucially for AI, it is rich where other graphs are thin: it “contains an abundance of ‘indications’, ‘contradictions’, and ‘off-label use’ drug-disease edges … and can support AI analyses of how drugs affect disease-associated networks.” They even “supplement PrimeKG’s graph structure with language descriptions of clinical guidelines to enable multimodal analyses.” A shared, open, continually updated map — the substrate the next model would learn on.

🧠 A foundation model that reasons over the graph

That model arrived. Huang et al. (Nature Medicine, 2024) named the ceiling of earlier tools plainly: “the clinical utility of drug-repurposing artificial intelligence (AI) models remains limited because these models focus narrowly on diseases for which some drugs already exist.” Their answer, TxGNN, is “a graph foundation model for zero-shot drug repurposing, identifying therapeutic candidates even for diseases with limited treatment options or no existing drugs.” Trained on a medical knowledge graph, it ranks drugs as indications and contraindications for 17,080 diseases and “improves prediction accuracy for indications by 49.2% and contraindications by 35.1% under stringent zero-shot evaluation.” And it doesn’t just answer — it explains: its “Explainer module offers transparent insights into multi-hop medical knowledge paths that form TxGNN’s predictive rationales,” with the reassuring result that “many of TxGNN’s new predictions align well with off-label prescriptions that clinicians previously made in a large healthcare system.” A model that can reason toward a disease with no treatment — and show its work.

🧭 The synthesis: graphs are the universal language

Step back, and a review by Li, Huang & Zitnik (Nature Biomedical Engineering, 2022) names why this whole family of methods keeps working: “networks—or graphs—are universal descriptors of systems of interacting elements.” Molecular interactions, signalling pathways, disease co-morbidities, whole healthcare systems — all are graphs, and the authors “posit that representation learning can realize principles of network medicine.” The horizon they sketch is broad and, tellingly, spans the rest of this digest’s beats: “the identification of genetic variants underlying complex traits, the disentanglement of single-cell behaviours and their effects on health, the assistance of patients in diagnosis and treatment, and the development of safe and effective medicines.” One representation, many biologies.

🧬 Why it’s our kind of problem

This is a different lens from the rest of the week, and that’s the point. Structure and sequence models ask “what is this molecule and how does it move”; knowledge graphs ask “how is everything connected” — and the two are complementary, as COVID network medicine showed when 76 of 77 hits acted through connections a docking model would never see. The reasoning style here is close to the lab’s heart: our Research Navigator is about reasoning over biomedical knowledge, and TxGNN’s multi-hop interpretable rationales are exactly the transparent, cite-your-path reasoning an AI co-scientist needs to be trusted. The ethos matches too — Hetionet, Decagon and PrimeKG are all open graphs and code, the public yardsticks everyone is measured on, in the same publish-the-model-and-the-data spirit behind the BioImage Model Zoo and BioEngine. And it’s a rung toward the virtual cell, which will need both lenses at once: the mechanistic physics of molecules and a knowledge-graph scaffold of how genes, drugs and diseases relate. A repurposing model you can query with an explanation attached, built on an open graph of biology: that’s AI for life science reasoning about the whole web, not just one thread of it.

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.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.