Lab Newsletter — September 6, 2026: The Wiring Diagram of a Cell

Sun, 06 Sep 2026 03:00:22 +0000

What makes a neuron a neuron and not a liver cell? It isn’t a single gene — both carry the same genome. It’s a circuit: a web of transcription factors switching one another on and off, holding the cell in one stable state rather than another. This week we’ve told the cell’s story in layers (multi-omics), in motion (RNA velocity), and in parts (protein design). Today’s is about the wiring — the causal diagram of who regulates whom. Reconstructing that diagram from data is gene regulatory network (GRN) inference, and as the GENIE3 authors put it plainly, it remains “one of the pressing open problems of computational systems biology”: the elucidation of network topology “using high throughput genomic data.” It’s one of the oldest hard problems in the field — and one where AI has quietly, steadily made headway.

🌲 Turning a network into regression

The elegant move that still powers much of the field came from GENIE3 (Huynh-Thu … Geurts, PLoS ONE, 2010), “a new algorithm for the inference of GRNs that was best performer in the DREAM4 In Silico Multifactorial challenge.” Instead of tackling the whole tangle at once, GENIE3 “decomposes the prediction of a regulatory network between p genes into p different regression problems” — for each gene, predict its expression from all the others, and read off which predictors mattered: “the importance of an input gene in the prediction of the target gene expression pattern is taken as an indication of a putative regulatory link.” Using random-forest ensembles, it “doesn’t make any assumption about the nature of gene regulation, can deal with combinatorial and non-linear interactions, produces directed GRNs, and is fast and scalable.” Sixteen years on, its engine still beats inside modern single-cell pipelines.

Inference is easy to do and hard to trust — anyone can output a network; is it right? The DREAM5 consortium (Marbach … Stolovitzky, Nature Methods, 2012) made this rigorous, opening on the same note of humility: “reconstructing gene regulatory networks from high-throughput data is a long-standing challenge.” They ran “a comprehensive blind assessment of over 30 network inference methods” across bacteria, yeast, and simulated data, and the headline result is one this digest keeps rediscovering: “no single inference method performs optimally across all data sets.” The fix was a crowd: “integration of predictions from multiple inference methods shows robust and high performance across diverse data sets.” Best of all, they didn’t stop at scores — they “experimentally tested 53 previously unobserved regulatory interactions in E. coli, of which 23 (43%) were supported,” establishing “community-based methods as a powerful and robust tool” for reading the circuit.

🔬 Down to single cells

Bulk expression averages over thousands of cells, blurring exactly the state-specific wiring you most want to see. SCENIC (Aibar … Aerts, Nature Methods, 2017) brought inference to the single-cell era as “a computational method for simultaneous gene regulatory network reconstruction and cell-state identification from single-cell RNA-seq data.” Its clever guardrail is biology: rather than trust correlations alone, it “exploit[s] cis-regulatory analysis … to guide the identification of transcription factors and cell states” — keeping only edges where the target gene actually carries the regulator’s DNA binding motif. The result, the authors report, “provides critical biological insights into the mechanisms driving cellular heterogeneity” — the network and the cell types falling out of the same analysis. (Under the hood, SCENIC’s inference engine is GENIE3’s descendant.)

📏 How good, really?

With a crowd of single-cell methods available, the field again needed a referee, and built a careful one. BEELINE (Pratapa … Murali, Nature Methods, 2020) delivered “a systematic evaluation of state-of-the-art algorithms for inferring gene regulatory networks from single-cell transcriptional data,” testing against synthetic networks, curated Boolean models, and real datasets. The verdict is bracingly honest — “the area under the precision-recall curve and early precision of the algorithms are moderate” — and the practical guidance is specific: “techniques that do not require pseudotime-ordered cells are generally more accurate.” It’s the same prove-it discipline we admire elsewhere — a benchmark whose job is to tell you how far there still is to go, built explicitly so it “will aid the development of gene regulatory network inference algorithms.”

🎛️ From diagram to simulator

Here’s the turn that makes all of this matter: a good GRN isn’t just a picture — it’s something you can run. CellOracle (Kamimoto … Morris, Nature, 2023) uses “gene-regulatory networks inferred from single-cell multi-omics data to perform in silico transcription factor perturbations, simulating the consequent changes in cell identity using only unperturbed wild-type data.” Read that again: from wild-type data alone, knock out a transcription factor on a computer and predict what the cell becomes. Applied to blood formation and zebrafish development, it “correctly model[s] reported changes in phenotype” — and, decisively, it “simulate[s] and experimentally validate[s] a previously unreported phenotype that results from the loss of noto, an established notochord regulator.” A prediction made in silico, confirmed at the bench: exactly the loop a self-driving lab is built to close.

🧬 The enhancer layer, and why it’s our problem

The frontier adds a layer we met earlier: chromatin. SCENIC+ (Bravo González-Blas … Aerts, Nature Methods, 2023) uses “joint profiling of chromatin accessibility and gene expression in individual cells … to decipher enhancer-driven gene regulatory networks,” predicting “genomic enhancers along with candidate upstream transcription factors (TFs)” and linking them “to candidate target genes” — wiring that runs TF→enhancer→gene, the way regulation actually works. It’s a fitting capstone to the week: Friday we designed the parts; today we map the circuit those parts run in. Reading a cell (function), writing its molecules (design), and wiring its logic (today) are three faces of one systems view — and a serious virtual cell or Human Cell Simulator needs all three. The causal diagram is the piece that turns description into simulation — knock out a gene in software, and predict the cell. And the way these tools travel is our ethos exactly: GENIE3, SCENIC, SCENIC+ and CellOracle are open source, DREAM5 and BEELINE are the public yardsticks — the same publish-the-model-and-the-test spirit behind the BioImage Model Zoo and BioEngine. Learn a cell’s wiring, and you can start asking what happens when you rewire 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.