One such effort, described in the same issue of the journal, is the Bits to Binders competition, in which Kosonocky, Abel and colleagues evaluated roughly 12,000 designed proteins as recognition domains for chimeric antigen receptor, or CAR, T cells. The scale of the undertaking is remarkable: twenty-eight teams submitted designs generated using diverse computational workflows, and each design was inserted into a common CAR scaffold for evaluation in primary human T cells. This design choice matters enormously, because it isolates the variable of interest, the binder itself, from the many other components that determine whether a cell therapy works. By holding the scaffold constant, the organizers could ask a deceptively simple question with profound implications: which AI-generated proteins can actually redirect a living immune cell to kill a cancer cell?

The molecular players at the center of the competition are minibinders, highly compact proteins that generative AI methods produce with increasing reliability. Unlike the antibody-derived single-chain variable fragments, or scFvs, conventionally used for antigen recognition in CAR T cells, minibinders typically comprise alpha helices and beta sheets arranged in small structures of roughly 65 to 80 amino acids, with well-folded architectures and programmable binding interfaces. While antibodies have historically been the dominant format for therapeutic protein binders, minibinders have already demonstrated heterogeneous uses in the academic literature, ranging from biosensors to viral neutralization, including picomolar inhibitors of SARS-CoV-2 designed de novo, to recognition domains in cell therapies targeting antigens in glioblastoma. Their ultimate therapeutic efficacy, however, has yet to be determined, which is precisely why a functional benchmark matters.

Earlier candidates from de novo protein design were largely assessed on biophysical parameters such as predicted stability, solubility, and affinity for a chosen target. Community competitions, including challenges supported by the company Adaptyv Bio, have begun to compare design workflows through standardized measurements of protein expression and binding affinity. But as the field moves toward repurposing these binders as therapeutic biologics, such criteria become necessary yet insufficient, because each therapeutic modality carries additional constraints. Extending the benchmarking framework to therapeutic applications requires tests that establish whether molecular recognition actually translates into the intended biological response. Bits to Binders was built to answer exactly that question, establishing an open community effort to assess the functional potential of AI-designed minibinders inside CAR T cells, the context in which they would ultimately need to perform.

The biological target chosen for the competition was CD20, a surface protein expressed by normal B cells and many B-cell malignancies. CARs work by repurposing antigen-binding domains to focus T cells toward specific target antigens on tumor cell surfaces, and the competition used the well-characterized antibody rituximab as a benchmark scFv alongside the de novo designs. Productive CAR function, however, demands far more than binding. A receptor must be expressed on the T cell surface, its binding epitope must be accessible on the target cell, and it must trigger appropriate signaling. Recent characterization of de novo CARs identified signaling without a target, inaccessible binding epitopes, and off-target activation as distinct constraints on efficacy that go well beyond simple binding, underscoring the importance of evaluating designed binders in their intended cellular context rather than in a test tube.

To capture these complexities at scale, the organizers used each binder-encoding sequence as a molecular barcode to track its corresponding CAR T cells in a pooled proliferation screen. After growth, 6,811 designs met the sequencing recovery threshold, and 707 of those showed significant enrichment in cultures containing CD20-positive lymphoma cells relative to cultures without them. The funnel then narrowed sharply. Individual testing of ten leading designs confirmed increased proliferation, expansion, and cytokine production for seven compared with unmodified T cells. Four showed significant selective killing in comparisons involving CD20-positive and CD20-negative cell lines. The designs generally underperformed the antibody-derived CAR control, indicating that further optimization remains necessary before AI-generated minibinders can match the refined binding domains that decades of antibody engineering have produced.

Perhaps the most sobering finding came from the biochemical side of the validation. Only three of the designs showed validated binding in biochemical assays, whose conditions differed from the cellular context of the proliferation screen. That disconnect between cell-based activity and cell-free binding measurements is itself an important result, because it suggests that computational predictions of affinity do not reliably predict function inside a living cell. Nevertheless, the retrospective analyses of sequences that failed at each step of the pipeline nominated concrete criteria for prospective testing, including screening DNA sequences for high GC content and repetitive elements, and limiting excessive enrichment of lysine and glutamate residues in helical structures. These are exactly the kind of actionable design rules that open competitions are uniquely positioned to generate, since no single laboratory could test such a diverse portfolio of approaches on its own.

The commentary authors are careful to note several caveats that signal opportunities to strengthen future competitions. The pooled screen assessed CAR behavior with or without lymphoma cells, but additional controls, such as isogenic knockout cell lines, would be needed to specifically tie activity to CD20 rather than to some other feature of the tumor cells. Testing a single receptor scaffold and a single T-cell donor also limits how far the sequence features identified in the binders can be generalized. Moreover, differences in the designed sequences reflect distinct workflows, including sequence selection and refinement steps, which makes it difficult to attribute performance to any individual algorithm. Future competitions that expand experimental validation and invite submissions annotated with the specific pipeline components being tested may better isolate the principles that separate efficacious therapeutic designs from expensive failures.

Looking forward, sustaining open competitions of this kind will require durable support for experimental testing, independent assessment, and shared data resources. The history of CASP offers a cautionary tale: despite decades of public funding, the competition faced lapses in support and secured a one-time private gift only in 2025 to keep running. As emerging competitions with more diverse endpoints, including Bits to Binders, expand the scope and promise of AI-enabled community benchmarking, they will similarly require sustained investment in the infrastructure necessary to realize biological breakthroughs. The stakes are considerable. If generative AI can reliably produce functional therapeutic binders, the design-build-test cycle for cell therapies could compress from years to months. Bits to Binders has now shown, with unusual transparency, both how far the technology has come and how much distance remains between a computationally elegant protein and a medicine that works in a patient.

Subject of Research: Benchmarking AI-designed minibinders as functional recognition domains in CAR T cell therapy through an open competition

Article Title: Crowdsourcing functional cell therapy binders from generative artificial intelligence

Article References: Dozic, A. V., & Lareau, C. A. (2026). Crowdsourcing functional cell therapy binders from generative artificial intelligence. Molecular Systems Biology. https://doi.org/10.1038/s44320-026-00249-y

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00249-y

Keywords: generative AI, protein design, minibinders, CAR T cells, Bits to Binders, CASP, CD20, immunotherapy, protein engineering, benchmarking, Molecular Systems Biology, computational biology

The research team, led by scientists at the Olivia Newton-John Cancer Research Institute, the Walter and Eliza Hall Institute, and collaborators in Australia and France, tackled a problem that has long frustrated the field. When a pathologist examines a needle biopsy, how much of the tumor’s clonal complexity is actually being seen? When circulating tumor DNA is detected in a patient’s plasma, which clones are shedding that DNA and which are hiding? Previous comparisons of the two approaches often focused on a handful of well-known driver mutations such as TP53 or KRAS, rather than the full landscape of clonal diversity. To go deeper, the team turned to a powerful lineage-tracing technology known as genetic barcoding.

The technique works like this: cancer cells are infected with a lentiviral library containing thousands of unique DNA sequences, each of which integrates permanently into the genome of a single cell and is inherited by all of its descendants. Every clone in the resulting tumor therefore carries a molecular name tag that can be recovered by PCR amplification and sequencing. In the study, the researchers applied this strategy to six preclinical breast cancer models, comprising two established cell lines, MDA-MB-231 and MDA-MB-468, and four patient-derived xenografts, or PDXs, established from drug-naive triple-negative breast cancer patients. Because the barcodes provide an unbiased record of clonal identity independent of any particular mutation, the researchers could track exactly which clones ended up in which sample.

The first major finding concerned geography. Using a stochastic agent-based simulation of three-dimensional tumor growth, adapted from a previously published model, the team predicted that clonal density would be higher in the center of a tumor than at its periphery. When they then dissected real barcoded tumors grown in mice into central and peripheral pieces and sequenced the barcode repertoire of each, the prediction held. In the MDA-MB-231 model, the tumor center contained on average roughly three times more distinct barcodes than peripheral pieces, and across all six models the center harbored a substantially larger share of the total clonal diversity. In one cell line model the center accounted for 66 percent of all detected barcodes, and in the other, 77 percent. The pattern was reproduced in all four PDX models, with the intriguing exception that a highly necrotic PDX showed a smaller center-periphery difference, presumably because cell death had erased many barcodes from the core.

This spatial bias has direct clinical implications, because radiologists typically aim needles toward the center of a lesion. The barcoding experiments showed that a needle biopsy reaching the tumor core captures a large fraction of the tumor’s biomass, with clones detected in needle samples representing roughly 80 to 90 percent of the total tumor reads. However, the content of any single needle varied considerably depending on the direction from which it was inserted. Several dominant clones present in the primary tumor were missed entirely by some needle tracks, and biopsies taken from opposite sides of the same tumor contained non-overlapping sets of barcodes. Minor clones, in particular, were frequently absent from needle samples, a sobering result given that low-frequency clones can drive metastasis and drug resistance.

The team then turned to the liquid side of the comparison. Blood was collected from tumor-bearing mice at three time points, corresponding to tumor volumes of 100, 300, and 800 cubic millimeters, and cell-free DNA was extracted from the plasma before the barcodes were amplified and sequenced. Barcode recovery in cfDNA proved strikingly variable. Some models, such as PDX-4295 and CRCM412, shed detectable DNA early and reliably, while others yielded little or nothing even when tumors were large. Recovery also increased with tumor burden, and within a given model the same clones could appear and disappear in the plasma of the same mouse over time. When the researchers implanted identical clonal pools into multiple recipient mice, the clonal composition of the resulting sister tumors was well conserved, yet barcode recovery still fluctuated between animals, suggesting that both intrinsic clone properties and stochastic factors influence how much DNA a tumor releases into the bloodstream.

When cfDNA barcodes were detected, they told a meaningful story. The plasma repertoire at the final time point broadly reflected the clonal makeup of the primary tumor, and the clones represented in cfDNA accounted for up to 80 percent of tumor biomass in the best-shedding model. Notably, cfDNA also served as a reasonable surrogate for the clonal composition of lung metastases, hinting that DNA in the blood may report on disease beyond the primary site. But the method had clear blind spots: only a small percentage of the tumor’s total barcodes appeared in the plasma in most models, and some dominant clones in the primary tumor were never detected in cfDNA at all.

The head-to-head comparison revealed that the two biopsy types are complementary rather than interchangeable. Deep needle biopsies generally captured a larger share of primary tumor barcodes and higher clonal diversity than cfDNA, except in the CRCM412 model, where cfDNA outperformed both deep and shallow needle sampling. Three distinct patterns emerged across individual mice: sometimes cfDNA better recapitulated tumor heterogeneity, sometimes the needle did, and sometimes both performed comparably. Crucially, combining the two methods significantly increased the percentage of tumor biomass captured compared with either alone in several models. In the MDA-MB-231 model, the researchers went one step further and retrieved circulating tumor cells from the blood; adding their barcodes to the combined needle-plus-cfDNA dataset pushed the captured biomass above 80 percent.

The authors are careful to note the limitations of their system. Xenograft experiments unfold over weeks rather than the years over which human tumors evolve, and barcode detection, which relies on amplifying tags shorter than 100 base pairs, may underestimate the cfDNA fragments detectable in clinical assays that query multiple genomic loci. Blood volumes in mice are also far smaller than those drawn from patients. Nevertheless, the barcoded models offer something patient samples cannot: a complete, ground-truth map of clonal composition against which every sampling method can be scored, tracked longitudinally, and manipulated experimentally without the confounding effects of treatment.

The broader takeaway for oncology is pragmatic. Solid biopsies deliver intact cells that support histology, receptor-status testing, single-cell omics, and assessment of tumor-infiltrating lymphocytes, while liquid biopsies offer a minimally invasive, repeatable readout that can integrate heterogeneity across multiple lesions that no needle could reach. This study provides quantitative backing for the growing consensus, supported by earlier analyses of hundreds of patient samples, that the two approaches should coexist in clinical practice. As barcoded preclinical models are combined with genomic and spatial transcriptomic profiling, researchers hope to identify the properties that make some clones prolific DNA shedders and others invisible, knowledge that could sharpen the interpretation of liquid biopsy results, flag false negatives, and ultimately guide more rational personalized therapies for breast cancer patients.

Subject of Research: Using genetic barcoding to quantify how well solid and liquid biopsies capture intra-tumoral heterogeneity in breast cancer models

Article Title: Genetic barcoding uncovers the clonal makeup of solid and liquid biopsies and their ability to capture intra-tumoral heterogeneity

Article References: Serrano, A., Weber, T. S., Berthelet, J., Ftouni, S., El-Saafin, F., Lee, S., Lim, E., Charafe-Jauffret, E., Ginestier, C., Williams, D., Hollande, F., Yeo, B., Dawson, S.-J., Naik, S. H., & Merino, D. (2026). Genetic barcoding uncovers the clonal makeup of solid and liquid biopsies and their ability to capture intra-tumoral heterogeneity. Molecular Systems Biology, 22(5), 659-684. https://doi.org/10.1038/s44320-026-00194-w

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00194-w

Keywords: breast cancer, genetic barcoding, intratumoral heterogeneity, liquid biopsy, cell-free DNA, needle biopsy, patient-derived xenografts, circulating tumor DNA, clonal evolution, tumor sampling, precision oncology, Molecular Systems Biology

The core insight behind the approach is disarmingly simple. AlphaFold3 can model the structure of protein complexes, and researchers have increasingly used it to test whether two proteins that have never been shown to interact might nevertheless form a physical pair. The obvious way to apply this to an entire organism is to run the algorithm on every possible pair of proteins one at a time. But the arithmetic quickly becomes brutal. The number of possible pairwise interactions scales quadratically with the size of the proteome, so even a modest bacterium generates tens of thousands of candidate pairs, while the human proteome, with roughly 20,000 proteins, yields around 200 million. Running that many individual AlphaFold jobs is, as the authors of the accompanying commentary note, computationally intractable with realistic resources.

Todor and colleagues sidestepped the problem by refusing to play the game one pair at a time. Instead of enumerating every possible interaction, they grouped proteins into random pools, generally containing between 10 and 25 proteins, and submitted each pool to AlphaFold3’s free online server as a single job. Because AlphaFold3 models all proteins in an input simultaneously, each pooled run produces an all-versus-all comparison within that pool. The pools were designed so that every pair of proteins in the proteome appears together in at least one pool, guaranteeing that the full set of pairwise interactions gets modeled once all the pools have been processed. The logic that makes this work is that true interactions are rare: out of the vast number of possible pairs, only a small fraction are real, so the extra proteins packed into a pool are unlikely to interfere with the evaluation of genuine interactions.

The efficiency gains are striking. Compared with the traditional pairwise approach, pooling reduced the total number of AlphaFold jobs by 100-fold and improved the runtime by twofold. More surprisingly, the pooled approach also improved the overall accuracy of identifying true interactions. For a set of seven proteins, the enumerated pairwise strategy would require 21 separate AlphaFold runs, while the pooling strategy accomplishes the same coverage in just three. In a field where compute budgets routinely dictate the scope of a study, a method that is simultaneously cheaper, faster, and more accurate is a rare and welcome combination.

The organism chosen for the demonstration, Mycoplasma genitalium, is a free-living bacterium with one of the smallest known genomes, containing roughly 475 proteins and therefore 113,050 possible pairwise interactions. That scale is manageable enough to serve as a proof of principle, but it also highlights the limits of the current achievement. The human proteome’s roughly 200 million possible pairs remain far beyond what pooling alone can conquer. Other research groups have attacked the scaling problem from a different direction, prioritizing pairs of proteins that are most likely to interact based on independent evidence. Burke and colleagues, for example, produced AlphaFold2 models of confident protein pairs drawn from publicly available human protein interaction databases such as hu.MAP 2.0 and HuRI, while other teams have applied related strategies to core eukaryotic complexes and to systematic human interactome computations.

Prioritization, however, has an inherent dependency: it requires experimental interaction data to exist for the species being studied, or at least for a close relative. Non-model organisms, which include most of the bacterial and archaeal diversity on the planet, often lack such data entirely. This is precisely where the pooling approach shines, since it needs no prior knowledge to get started. Importantly, the two strategies are not in conflict. When experimental data are available, pooling and prioritization could in principle be combined in future efforts, using evidence-based filters to narrow the candidate space and pooled AlphaFold3 screening to evaluate what remains. For understudied organisms, pooling stands as a valuable new addition to the computational toolkit.

Why pooling improves accuracy remains an open question, and the authors have offered several hypotheses. One appealing idea involves competition among interactions. Because the pooled approach evaluates many potential interactions simultaneously, only the most confident interaction at any given protein interface will be selected in the final model. This competitive filtering could suppress false positives, particularly for promiscuous proteins that have a tendency to produce spurious predictions when considered in isolation. A deeper understanding of the mechanism behind the accuracy boost would be more than an academic curiosity; it would allow researchers to engineer protein pools deliberately, optimizing their composition to maximize predictive performance rather than relying on random assignment.

That same hypothesis, however, exposes a potential weakness. If AlphaFold3 prioritizes some interactions based on spatial constraints, then mutually exclusive interactions, in which two different proteins bind a third protein at the same interface, may cause one partner to be systematically favored over the other. Such exclusivity is common in nature and often underlies molecular functions that switch depending on cellular context. The severity of the problem depends on how many mutually exclusive interactions exist and on the size of the pools. With the current setup of 10 to 25 proteins per pool, the issue is likely to be rare, but if completeness becomes the explicit goal, or if pool sizes are increased to gain further efficiency, it will need to be addressed head-on.

There are also limitations inherited from AlphaFold3 itself. The algorithm depends on multiple sequence alignments to identify co-evolving pairs of amino acids, both within a single protein and between proteins in a complex, and it uses those co-evolutionary signals as distance constraints when building structural models. Proteins without sufficiently deep alignments suffer greatly in accuracy. In Mycoplasma genitalium, around 40 proteins, roughly 8 percent of the proteome, have limited sequence homology and appear to be found only in this species. As a result, nearly 18,000 protein pairs cannot be evaluated by this approach at all. That gap is more than a technical footnote, because species-specific proteins and their interactions may be responsible for the unique biology of the organism. Context-dependent interactions, such as those that form only under particular cellular conditions or in specific cell types, are similarly difficult for the method to discern.

Looking ahead, pooled screening offers an intriguing path toward detecting higher-order interactions, such as trimers and tetramers, which are often more biologically relevant than simple pairwise contacts. The obstacle is probabilistic: purely random pools make it unlikely that three or more proteins that genuinely interact in the cell will land in the same pool. Capturing all pairwise interactions through pooling is likely to be effective, but exhaustively screening all possible three-way combinations, roughly 18 million in Mycoplasma genitalium alone, will remain out of reach for quite a while unless pools are chosen intelligently rather than at random. Even so, the trajectory is clear. Todor and colleagues have brought the field measurably closer to the long-sought goal of complete, all-versus-all computational screening of protein interaction landscapes, for model organisms and neglected ones alike. The efficiency gains and the unexpected accuracy boost suggest that pooled AlphaFold screening will become an indispensable tool in a toolkit that is expanding at a remarkable pace.

Subject of Research: Proteome-wide prediction of protein-protein interactions using pooled AlphaFold3 screening

Article Title: Proteome-wide AlphaFold pool party

Article References: Drew, K. (2026). Proteome-wide AlphaFold pool party. Molecular Systems Biology, 22(4), 477-479. https://doi.org/10.1038/s44320-026-00198-6

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00198-6

Keywords: AlphaFold3, protein-protein interactions, Mycoplasma genitalium, structural biology, computational biology, proteome, deep learning, protein complexes, interactome, multiple sequence alignment, molecular systems biology, Proteome-wide

The concept rests on a principle borrowed from evolutionary ecology. When cancer cells evolve resistance to a first-line drug, the molecular machinery they deploy often carries hidden costs or collateral vulnerabilities. A collateral sensitivity arises when adaptation to one drug increases susceptibility to another. The researchers hypothesized that if a second drug could exploit such a vulnerability, then applying selective pressure with chemotherapy would actively steer the tumor population into a corner: cells that resist chemotherapy become exquisitely sensitive to the second agent, leaving them nowhere evolutionarily to hide. This is the double-bind, a strategy first theorized in mathematical oncology but rarely demonstrated with a concrete drug pair and a validated treatment schedule.

To find such a pair, the team, led by Rena Emond of City of Hope and Jeffrey West of the Moffitt Cancer Center, worked with two ER+ breast cancer cell lines, MCF-7 and T-47D. Over six to eight months, they evolved sublines resistant to two common chemotherapies, paclitaxel and doxorubicin, by exposing cells to weekly pulses of each drug. They then screened thirty-seven candidate compounds, spanning FDA-approved oncology drugs, agents tested in clinical trials, and inhibitors of pathways linked to resistance. The screen measured whether each candidate preferentially killed resistant cells relative to their sensitive parents. Disulfiram, an inhibitor of aldehyde dehydrogenase, stood out: across all four chemotherapy-resistant line pairs, it showed dose-dependent toxicity that spared the sensitive parental cells, exactly the asymmetry a double-bind requires.

With disulfiram identified, the next question was whether combining it with chemotherapy produced true synergy, mere additivity, or antagonism. The researchers ran full checkerboard dose-response assays on both sensitive and resistant lines and analyzed the results with the MuSyC framework, which separates synergy into two components: synergistic potency, a shift in the half-maximal effective concentration, and synergistic efficacy, an increase in the maximum killing achieved at high doses of both drugs. The results were nuanced. In treatment-naive cells, the combination was largely neutral, meaning the two drugs acted additively rather than synergistically. In most resistant lines the additivity held, with one notable exception: doxorubicin-resistant MCF-7 cells displayed genuine synergistic potency with the combination. Importantly, antagonism was rarely observed, and the high-dose combination achieved maximal killing across every cell line and resistance background tested.

One curious finding emerged from the dose-response curves. Disulfiram alone behaved non-monotonically: intermediate doses killed more cells than the highest doses, producing a paradoxical dip in efficacy at high concentrations. When chemotherapy was added, this instability vanished, and the combination curve became smoothly dose-dependent. The authors suggest chemotherapy stabilizes disulfiram’s drug effect, an observation that underscores how combination contexts can rescue drugs whose monotherapy behavior is unpredictable. The team also notes that even additive combinations can be clinically valuable, a point increasingly recognized as trials of supposedly synergistic combinations often deliver benefits explainable by simple independent drug action.

Drug-drug interactions, however, tell only half the story. Tumors are ecosystems in which sensitive and resistant cells compete, cooperate, and modulate each other’s growth. To capture these dynamics, the researchers turned to the evolutionary game assay, a technique that quantifies frequency-dependent growth rates. They fluorescently labeled sensitive cells with one protein and resistant cells with another, then grew them as three-dimensional spheroids at initial resistant fractions of zero, twenty-five, fifty, seventy-five, and one hundred percent. Under untreated conditions, sensitive cells grew faster across all mixtures, revealing a fitness cost to resistance. Under chemotherapy, resistant cells held their advantage as expected. Under disulfiram, the picture flipped: resistant cells lost fitness to a far greater degree than sensitive ones, confirming the collateral sensitivity in a competitive, three-dimensional setting that more closely mimics the tumor microenvironment than flat culture dishes.

From these experiments the team built a game-theoretic mathematical model. Each cell type’s growth rate is described as a linear function of the population’s composition, yielding a four-parameter payoff matrix. Two derived quantities, the invasion fitnesses of sensitive and resistant cells when rare, determine the selection dynamics: whether sensitive cells dominate, resistant cells take over, the two coexist, or the system is bistable. The model also defines an overall tumor fitness, the long-term growth rate of the whole spheroid once selection reaches equilibrium. The critical prediction was unambiguous: continuous combination therapy with chemotherapy and a high dose of disulfiram drove tumor fitness to its most negative value, meaning the entire population, sensitive and resistant alike, would shrink rather than merely shift in composition.

The model then confronted a practical clinical question: should the two drugs be given together, or sequentially? Because the double-bind depends on resistance to the first drug, one might reasonably argue for chemotherapy first and disulfiram later. The team simulated alternating schedules with fast weekly and slower biweekly switching, always beginning with chemotherapy since disulfiram performs best once resistance has been selected. Across all parameterizations, the simulations predicted that combination therapy would outperform every sequential regimen. The researchers then tested this prediction directly in their spheroid system, treating mixed populations with twenty-five nanomolar doxorubicin and five hundred nanomolar disulfiram either continuously or in alternating schedules. The experiments confirmed the mathematics: combination treatment suppressed total cell count more effectively than any sequential strategy, across both cell lines and all initial resistant fractions. Interestingly, in MCF-7 cultures, leading a sequential schedule with disulfiram controlled growth better than leading with chemotherapy, but neither alternating approach matched the continuous combination.

The choice of disulfiram carries practical appeal beyond its evolutionary credentials. As an FDA-approved drug with decades of safety data, it can be repurposed without the lengthy development timeline of a novel molecule. Clinical trials have already explored disulfiram combinations in metastatic pancreatic cancer, treatment-refractory sarcoma, and recurrent glioblastoma, though the glioblastoma trial showed no significant survival benefit. Preclinical work suggests multiple mechanisms by which disulfiram might strike at chemo-resistant cells: it inhibits PI3K signaling, blocks ABC drug efflux pumps that drive multidrug resistance, suppresses NF-kB and epithelial-mesenchymal transition, undermines breast cancer stem cell self-renewal, and its copper-dependent metabolite induces p53-mediated apoptotic cell death. The present study did not pinpoint which of these mechanisms underlies the observed collateral sensitivity, and the authors flag mechanistic dissection as a priority for future work.

The researchers are careful about scope. These results come from cell lines and spheroids, not patients, and translating the double-bind into the clinic will require in vivo studies to establish optimal dosing, timing, and toxicity, particularly since disulfiram is unlikely to be administered alone and may behave differently in copper-containing combinations. Still, the study delivers something rarer than another drug pair: a validated, quantitative framework that integrates drug screening, evolutionary game assays, and mathematical modeling to predict which treatment schedule best suppresses resistance before a single patient is treated. If the evolutionary double-bind holds up in animal models and trials, the strategy could reframe how oncologists think about drug resistance, not as an inevitable failure of chemotherapy, but as a predictable vulnerability waiting to be exploited.

Subject of Research: Evolutionary double-bind combination therapy with disulfiram and chemotherapy to suppress drug resistance in ER+ breast cancer

Article Title: A novel combination therapy for ER+ breast cancer suppresses drug resistance via an evolutionary double-bind

Article References: Emond, R., West, J., Grolmusz, V. K., Cosgrove, P. A., Nath, A., Anderson, A. R. A., & Bild, A. H. (2026). A novel combination therapy for ER+ breast cancer suppresses drug resistance via an evolutionary double-bind. Molecular Systems Biology, 22(7), 1070-1096. https://doi.org/10.1038/s44320-026-00191-z

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00191-z

Keywords: ER+ breast cancer, drug resistance, disulfiram, chemotherapy, collateral sensitivity, evolutionary game theory, combination therapy, drug repurposing, mathematical modeling, spheroids, Molecular Systems Biology, cancer evolution

The core logic of MIDAS is disarmingly simple. Instead of building plasmids, researchers design overlapping PCR fragments that together encode a complete transcription unit, including a promoter, the variant coding sequence, and a polyadenylation signal. Mutations are introduced through primers at defined target regions, and a secondary overlap-extension PCR stitches the fragments into full-length genes. Each PCR product encodes one defined variant and is transfected into one well of a multiwell plate, so the association between sequence and measured function is preserved by well position alone. Because no ligation, transformation, bacterial culture, plasmid purification, or sequence verification is required, generating 384 variants takes roughly four hours of hands-on time and about 2,000 dollars in reagents, compared with an estimated 192 hours and 20,000 dollars for conventional cloning of the same number of constructs.

The platform comes in several flavors that expand its reach beyond single-site substitutions. Polytemplated monofocal MIDAS, or MIDAS-PM, uses multiple plasmid templates to generate arrays of mutants at one location. Polytemplated polyfocal MIDAS, or MIDAS-PP, extends this to combinations of changes at multiple sites by mixing primary PCR products combinatorially. Most notably, the team developed monotemplated protocols, MIDAS-MM and MIDAS-MP, which work from a single plasmid template. The trick lies in adding unique DNA tags to the 5-prime ends of the outermost primers, so that secondary PCR primers recognize only the assembled primary products and never re-amplify the parental plasmid. Sequencing confirmed that this nested-primer design eliminates template contamination, meaning researchers can begin mutagenizing a protein of interest immediately without first subcloning it into multiple backbone plasmids.

To demonstrate the method, the team took on an ambitious target: a bioluminescent indicator for acetylcholine, a neurotransmitter central to arousal, attention, learning, and memory, and one whose signaling declines early in Alzheimer’s, Parkinson’s, and Lewy body dementias. Fluorescent acetylcholine sensors exist, but they require implanted optical elements to deliver excitation light, which is impractical for moving organs outside the brain. A bioluminescent sensor, by contrast, generates its own light and could report cholinergic activity noninvasively anywhere in the body. The researchers built their prototype, ACh-NeuBI0.1, by inserting split NanoLuc luciferase fragments into OpuBC, a bacterial periplasmic binding protein previously engineered to bind acetylcholine over choline. The initial construct responded to acetylcholine with a modest 22 percent luminescence increase, a starting point that would demand extensive optimization.

MIDAS delivered that optimization in rapid iterative rounds. First, MIDAS-MP screened all 36 combinations of glycine linker lengths connecting the binding protein to the luciferase fragments, identifying a configuration that boosted responsiveness by roughly 50 percent. Next, MIDAS-MM tested all 20 amino acids at a single linker position, revealing that phenylalanine improved the response to about 80 percent. A third round tuned the energetics of luciferase fragment assembly by screening SmBiT variants, finding that a three-residue C-terminal truncation raised the response to approximately 2.5-fold. The team then fused the orange fluorescent protein mScarlet-I to the sensor, enabling resonance energy transfer that shifts a portion of the emission above 610 nanometers, wavelengths that penetrate tissue far more effectively than NanoLuc’s native blue light. At each step, plasmid-based validation confirmed that the PCR-transfected results were reliable.

The deepest round of engineering showcased MIDAS’s capacity for true combinatorial saturation mutagenesis. Guided by computational structure predictions from Chai-1, the researchers selected 16 sites around the acetylcholine-binding pocket and screened every possible amino acid at each, uncovering beneficial mutations at three second-shell positions. Because positions 560 and 610 sit close together in space, the team then used MIDAS-MP to test all 400 combinations of side chains at those two residues. Two winners emerged: variant QAE, named ACh-NeuBI1b, which achieved the highest affinity at 188 micromolar, and variant QED, named ACh-NeuBI1c, which suppressed baseline luminescence to deliver maximum contrast. Strikingly, ACh-NeuBI1c responded to 100 micromolar acetylcholine with a 640 percent signal increase, a 29-fold improvement over the original prototype.

The mathematics of why this matters are sobering. Seven of the nine beneficial amino acid changes discovered in the study required two or three nucleotide substitutions. Under error-prone PCR, a functional clone carrying two specific mutations appears roughly once in every 71 million clones, and some combinations once in 140 million, numbers far beyond what multiwell screening can reach. Deterministic PCR assembly sidesteps this barrier by constructing exactly the variants of interest, one per well, whether that means 20 samples for one site or 400 for two. The approach also outperforms degenerate oligonucleotide libraries, which require oversampling that balloons combinatorially: two sites of saturation mutagenesis would demand more than 10,000 screened samples, roughly 25 times what MIDAS requires.

The team applied the same logic to NanoLuc luciferase itself, performing saturation mutagenesis at all 25 residues lining or influencing the substrate-binding pocket, generating 500 variants in a single MIDAS-MM campaign. A D108N mutation improved photon production across three different substrates, while D108S, which requires two adjacent nucleotide changes and would be vanishingly rare in random libraries, selectively enhanced activity on the aqueous substrate FFz. Beyond engineering, the complete activity matrix enabled sequence-fitness analysis, mapping which active-site positions tolerate substitution and which are intolerable, and revealing substrate-specificity determinants whose locations often defied structural prediction. Because NanoLuc produces too few photons for fluorescence-activated cell sorting, the multiwell format proved essential, connecting MIDAS to the gold-standard analytical assays, from luminescence to absorbance to chromatography, that apply to most enzymes.

The biological payoff extended from cell culture into living animals. In primary cortical neurons, ACh-NeuBI1b and ACh-NeuBI1c achieved acetylcholine responses of up to 200-fold while maintaining selectivity over choline and other neurotransmitters. When expressed in mouse liver by hydrodynamic transfection, the optimized indicators detected intraperitoneally administered acetylcholine with signal increases of up to 173-fold, far exceeding the 30-fold response of the pre-optimization sensor, while an unregulated control reporter showed no response. The developers emphasize that improvements measured in mammalian cells, including affinity gains that did not reproduce when proteins were screened as bacterial periplasmic constructs, translated directly into in vivo performance, underscoring the value of screening in the environment where a protein must actually work.

Looking forward, the authors position MIDAS as a data engine for the machine-learning era of protein design. Computational models excel at proposing candidate sequences but cannot reliably identify the single best one, and they remain poorly trained on the dynamic conformational changes underlying enzyme catalysis and biosensor activation. By generating hundreds of precisely defined variants with high-quality functional measurements in mammalian cells within a day, MIDAS can supply exactly the kind of dense sequence-fitness datasets that such models need. The method blurs the traditional boundary between screening and analysis, unifying them in a single workflow that is roughly 48 times faster and ten times cheaper than cloning-based approaches. For laboratories optimizing biosensors, enzymes, or therapeutic proteins in mammalian systems, the bacterial incubator may no longer be the obligatory first stop.

Subject of Research: A cloning-free PCR-based platform for rapid protein engineering and sequence-fitness analysis in mammalian cells

Article Title: Fast analysis and engineering of protein function by microbe-independent deep assembly and screening

Article References: Wu, Y., Wang, P., Liu, L. X., Song, D., Qin, Q., Gao, C., Hageman, M., Kirkland, T. A., Su, Y., & Lin, M. Z. (2026). Fast analysis and engineering of protein function by microbe-independent deep assembly and screening. Molecular Systems Biology, 22(6), 1003-1034. https://doi.org/10.1038/s44320-026-00210-z

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00210-z

Keywords: protein engineering, MIDAS, PCR assembly, mammalian cells, directed evolution, bioluminescent indicator, acetylcholine, NanoLuc, saturation mutagenesis, sequence-fitness analysis, biosensors, Molecular Systems Biology

That blind spot matters more than many researchers realized. A cell’s fate is not determined solely by the genes it expresses; it is shaped by its position within a tissue, by the density of surrounding cells, by physical constraints, and by local signaling interactions. Conventional RNA velocity methods treat cells as independent entities floating in an abstract expression space, and the result is often inferred dynamics that contradict the actual architecture of the organ being studied. Trajectories may point across anatomical boundaries where no plausible transition exists, or flows may run in directions that make no sense given the known organization of, say, the cerebral cortex. Now a team reporting in Molecular Systems Biology has built a computational framework designed to fix exactly this problem, and in doing so has added something genuinely new: the ability to simulate how perturbing specific genes would redirect cellular dynamics within intact tissue.

The new method, called veloAgent, was developed by Raghavan, Yoon, Fonseca, Li, and Ding and described in a companion News & Views commentary by Jiahao Wu and Suoqin Jin of Wuhan University. At its core, veloAgent takes spliced and unspliced mRNA counts as input and passes them through a variational autoencoder, a type of deep generative neural network that compresses high-dimensional expression data into a compact latent representation. This compression serves two purposes. It reduces the noise and dimensionality that plague raw count data, and it produces a robust summary of each cell’s transcriptional state that downstream models can work with more reliably than with raw counts alone.

From that latent representation, a neural network constructed in conjunction with gene-gene interaction networks predicts kinetic parameters for every cell and every gene: the transcription rate alpha, the splicing rate beta, and the degradation rate gamma. These are the same mechanistic quantities that underlie classical RNA velocity theory, but here they are estimated with the benefit of regulatory structure encoded in known interactions between genes. The initial velocity computed from these parameters is then refined in two stages. The first stage applies velocity smoothing, aligning the directions of transcriptionally similar cells to enforce local coherence and suppress erratic, single-cell noise. The second stage is where veloAgent departs most sharply from its predecessors.

That second stage is an agent-based model, a simulation technique borrowed from fields like epidemiology and economics in which individual entities, or agents, interact according to local rules and collective behavior emerges from those interactions. In veloAgent, each cell is an agent embedded in its spatial microenvironment. The velocity assigned to each cell is adjusted through neighborhood-based weighting that incorporates spatial distance, local cell density, and expression similarity. In practical terms, a cell’s inferred trajectory is constrained by what its actual physical neighbors are doing. Rather than treating spatial information as an auxiliary feature bolted onto the analysis, veloAgent uses it as a primary organizing principle, allowing cell-state transitions to emerge from local interactions rather than being imposed from above. The refined velocity field then supports downstream analyses including spatio-temporal dynamics inference, projection of velocity fields onto low-dimensional spaces and spatial tissue maps, and in silico gene perturbation.

The perturbation capability deserves particular attention because it moves the framework from description toward prediction. Instead of simply altering observed expression values, veloAgent modulates the underlying kinetic parameters directly. Increasing or decreasing alpha for a chosen gene simulates transcriptional activation or repression at the level of the dynamical system itself, after which the induced velocity field is recomputed. The functional impact of each simulated perturbation is quantified by measuring how the directionality of cellular velocity vectors changes, particularly with respect to predefined target states or trajectories. Genes are then ranked by their ability to reorient cellular dynamics toward those targets, yielding a prioritized list of candidate regulators of cell-state transitions. Because the perturbations operate within the learned parameter space, large-scale screening can be performed efficiently without retraining the model, while retaining mechanistic interpretability that purely correlative approaches lack.

The authors validated veloAgent on spatial transcriptomics datasets from multiple platforms and across diverse tissue types. In structured tissues such as the brain, the framework produced trajectories that closely aligned with known anatomical organization, avoiding the inconsistent or anatomically implausible directions that earlier methods generated. In silico perturbation analysis identified key regulatory genes whose simulated modulation reversed progression toward terminal fiber-tract cell states, demonstrating that the model can pinpoint levers that genuinely redirect cellular dynamics. In a breast cancer dataset, veloAgent’s inferred transitions were more consistent with expected tumor progression patterns, correcting misdirected flows produced by competing methods, and further perturbation analysis revealed potential therapeutic targets capable of suppressing malignant progression.

The quantitative benchmarks reinforce the qualitative picture. veloAgent outperformed alternative methods in cross-boundary directionality, fate probability, and velocity confidence, the metrics that capture whether inferred trajectories respect tissue boundaries and how reliably they predict cell fates. Notably, the framework also improved temporal velocity estimation for standard single-cell RNA sequencing data even when no spatial information was available, suggesting that the agent-based refinement and gene-gene regulatory constraints confer benefits that extend beyond the spatial setting. For a field in which velocity estimates have often been fragile and method-dependent, that robustness is a meaningful advance.

The broader significance, as Wu and Jin emphasize in their commentary, is that cellular behavior cannot be understood outside its native environment. Transcriptional state, spatial structure, physical constraint, and local interaction all conspire to determine what a cell becomes. By modeling cells as interacting agents within tissue rather than as independent observations, veloAgent offers a more realistic representation of how cellular processes actually unfold, and it provides a mechanistically grounded platform for simulating interventions before any experiment is run in the laboratory.

Challenges remain, and the commentary is candid about them. Spatial transcriptomics data are frequently collected across multiple tissue sections and experimental conditions, creating batch-correction problems and complicating robust cross-sample comparison of cellular dynamics. Integrating additional modalities such as epigenomics and proteomics would give a fuller picture of the regulatory processes driving cell dynamics. Incorporating causal machine learning frameworks could sharpen both interpretability and predictive power, and directly modeling intercellular signaling, rather than approximating neighborhood effects under spatial constraints, would deepen the mechanistic account of how cells influence one another. Still, the trajectory of the field is clear. The long-term ambition, articulated in recent proposals for building virtual cells with artificial intelligence, is to move beyond descriptive analysis toward predictive models capable of simulating cellular behavior in silico, accelerating therapeutic target discovery and advancing precision medicine. veloAgent represents a concrete step along that path: a tool that not only decodes where cells in a tissue are going, but suggests how researchers might steer them somewhere else.

Subject of Research: Spatially resolved RNA velocity modeling and in silico perturbation of cell-state transitions in tissue

Article Title: Decoding and steering spatially resolved cellular dynamics

Article References: Decoding and steering spatially resolved cellular dynamics. (n.d.). https://doi.org/10.1038/s44320-026-00216-7

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00216-7

Keywords: RNA velocity, spatial transcriptomics, single-cell RNA sequencing, variational autoencoder, agent-based model, cell-state transitions, gene perturbation, computational biology, tumor progression, deep learning, gene regulation, Molecular Systems Biology

Protein-protein interactions are the machinery of cellular life. Enzymes pair with substrates, scaffolds anchor signalling cascades, and chaperones shepherd newly minted proteins into their correct folds. Cataloguing these interactions at scale has been a major goal of proteomics for two decades, but most existing methods carry significant limitations. Affinity purification-mass spectrometry requires targeting individual proteins and can lose weak or transient partners during sample preparation. Protein correlation profiling, which infers interactions from the co-migration of proteins through a separation column, works at proteome scale and avoids tagging, but it cannot confirm that two proteins bind directly. Cross-linking mass spectrometry solves the directness problem by covalently stitching together proteins that are in close physical proximity, but capturing a comprehensive interactome from intact tissue has proven exceptionally difficult because cross-linkers penetrate tissue poorly and cross-linked peptides are vanishingly rare in complex mixtures.

The Melbourne team attacked the problem with a dual strategy. In cell culture, they induced insulin resistance in differentiated C2C12 mouse myotubes by exposing them to the fatty acid palmitate for 24 hours, a manipulation that faithfully blunted insulin-stimulated phosphorylation of Akt and its substrate AS160, canonical markers of impaired insulin signalling. They then applied two orthogonal technologies. The first, protein correlation profiling mass spectrometry, involved lysing cells in gentle native buffer, separating intact protein complexes by blue native polyacrylamide gel electrophoresis, slicing each gel lane into thirty fractions, and analysing all 360 resulting fractions by data-independent acquisition mass spectrometry. Computational analysis with the CCProfiler algorithm mapped how thousands of protein complexes reorganise between healthy, insulin-resistant, and acutely insulin-stimulated states.

The second technology was quantitative cross-linking mass spectrometry using a membrane-permeable, phosphate-enrichable cross-linker called t-butyl-PhoX, combined with sixteen-plex tandem mass tag labelling for multiplexed relative quantification. After cross-linking, the researchers subjected samples to subcellular fractionation into cytoplasmic, membrane, nuclear, and cytoskeletal compartments, digested the proteins, depleted phosphopeptides, enriched PhoX-modified cross-linked peptides by immobilised metal-ion affinity chromatography, and then passed them through size-exclusion and high-pH reversed-phase chromatography before mass spectrometric analysis. This deep fractionation cascade was essential: cross-linked peptides are dramatically outnumbered by ordinary peptides, and without such enrichment they would be lost in the noise. Data were searched with the pLink2 engine against a skeletal muscle-specific mouse proteome database and filtered with xiFDR to stringent false discovery thresholds at the peptide, residue-pair, and protein-interaction levels.

The scale of the resulting datasets is striking. Across the cell-based experiments, the integrated workflows quantified more than 7,000 unique protein-protein interactions among 5,346 proteins. In the cross-linking analysis of myotubes, the team identified 11,295 cross-linked peptides, including 4,944 that linked two different proteins, corresponding to 3,757 unique protein pairs. Comparing the cross-link abundances between healthy and insulin-resistant cells revealed that 6,704 cross-linked peptides were significantly regulated, and crucially, 5,214 of those changed without any accompanying change in the abundance of either interacting partner — direct evidence that the interaction network itself, not merely protein expression, is being remodelled by the disease process.

Pathway analysis of the regulated cross-links pointed overwhelmingly toward one cellular compartment: the endoplasmic reticulum, the organelle where secreted and membrane proteins are folded, quality-controlled, and dispatched. Interactions among ER-resident chaperones and folding catalysts — including multiple protein disulfide isomerases, the calcium-binding chaperone calreticulin, the stress-induced chaperone HYOU1, and BiP, the master regulator of the unfolded protein response — were systematically perturbed. This converges with a substantial literature linking ER stress to insulin resistance, most famously the demonstration that ER stress activates the JNK kinase, which inhibits insulin receptor substrate-1 and thereby suppresses insulin signalling. What the new study adds is a systems-level picture of how the chaperone network’s physical architecture is dismantled in the insulin-resistant cell, suggesting the muscle loses capacity to resolve folding stress precisely when it is most needed.

One protein emerged as a central hub of this dysregulated network: PDIA6, an ER-resident protein disulfide isomerase that catalyses the formation and rearrangement of disulfide bonds, thereby tuning the oxidation state of cysteine residues in its client proteins. The team identified eighteen direct interactions with PDIA6, several of them not annotated in the latest human interactome databases, and all thirty-four cross-linked peptides bridging PDIA6 and BiP were downregulated with insulin resistance. Protein correlation profiling independently showed PDIA6 redistributing from its monomeric molecular weight into larger assemblies under insulin-resistant conditions. This echoes recent work showing that PDIA6 forms calcium-dependent biomolecular condensates in the ER lumen that anchor other folding chaperones under homeostatic conditions, dispersing throughout the ER during stress to increase folding capacity at the cost of substrate specificity.

To test whether these interactome changes matter functionally, the researchers overexpressed PDIA6 throughout the skeletal musculature of young mice using a muscle-tropic viral vector, then fed the animals a high-fat diet for twelve weeks to induce insulin resistance. Western blotting confirmed a five-fold overexpression. Contrary to the team’s hypothesis that boosting chaperone capacity would protect insulin sensitivity, the overexpressing mice showed subtly reduced insulin-stimulated glucose uptake in isolated soleus muscle and diminished phosphorylation of Akt and AS160. Cysteine redox proteomics using isobaric iodoTMT labelling then quantified 578 cysteine-containing peptides in the overexpressing muscles, revealing that the most significantly oxidised cysteines sat on proteins physically cross-linked to PDIA6, including the chaperones MANF, calreticulin, TXNDC5, and PDIA3, alongside calcium-handling proteins such as the calcium pump ATP2A1 and sarcalumenin.

Perhaps the most technically ambitious component of the study was the extension of quantitative cross-linking mass spectrometry to intact skeletal muscle tissue. The researchers excised extensor digitorum longus muscles from chow-fed and high-fat-diet-fed mice, incubated the freshly isolated tissues with the cross-linker in oxygenated physiological buffer for two hours, and pushed them through the same deep-fractionation pipeline. This yielded 8,327 cross-linked peptides spanning 3,541 unique protein pairs, of which 1,101 were significantly regulated by the high-fat diet. Regulated interactions clustered in glycolysis, calcium signalling, and contractile machinery, including interactions among the calcium release channel RYR1, calsequestrin, triadin, and SERCA1. A striking pattern of regulated intra-protein cross-links concentrated on the light meromyosin region of myosin heavy chain 7, consistent with conformational shifts between the energy-saving super-relaxed state and the disordered-relaxed state that have recently been implicated in diabetic muscle energetics.

Validation of the workflow came from mapping cross-links onto experimentally solved structures of the immunoproteasome, mitochondrial complex I, the ARP2/3 complex, and the actomyosin complex: 97.5 percent of mapped cross-links fell within the expected thirty-five angstrom distance constraint of the cross-linker, with the few violations occurring in regions known to flex during contraction. The authors are candid about limitations — cross-linker infusion into living animals remains impractical, the ex vivo cross-linking window may itself perturb the interactome, and the extensive fractionation could bias which proteins are captured. Nevertheless, the study establishes that quantitative interactome mapping is feasible in whole tissue and delivers biological insight unattainable from abundance measurements alone. By showing that the rewiring of protein interaction networks — and specifically the disulfide-isomerase activities of PDIA6 — reshapes cysteine oxidation and insulin sensitivity in muscle, the work opens a genuinely new dimension in the search for the molecular roots of metabolic disease.

Subject of Research: Quantitative mapping of protein-protein interaction remodelling in skeletal muscle insulin resistance

Article Title: Quantitative interactome mapping of skeletal muscle insulin resistance

Article References: Ng, Y.-K., Blazev, R., Wong, J. P. H., Scott, N. E., Molendijk, J., & Parker, B. L. (2026). Quantitative interactome mapping of skeletal muscle insulin resistance. Molecular Systems Biology, 22(9), 1430-1452. https://doi.org/10.1038/s44320-026-00224-7

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00224-7

Keywords: insulin resistance, skeletal muscle, protein-protein interactions, cross-linking mass spectrometry, proteomics, PDIA6, endoplasmic reticulum stress, diabetes, interactome mapping, cysteine oxidation, type 2 diabetes, molecular systems biology

The team combined a microfluidic device known as the dual input mother machine with time-lapse fluorescence microscopy and automated image analysis to follow thousands of individual cells through an abrupt switch from glucose-rich growth medium to a carbon-free version of the same medium. Each strain carried a transcriptional fluorescent reporter for one of 22 different promoters, spanning native genes and synthetic constitutive promoters chosen to cover a wide range of expression levels and regulatory behaviors. This multiplexed setup allowed the researchers to quantify, for every single cell, its size, growth rate and the rate at which it produced a fluorescent protein from each promoter, corrected for the confounding effects of photobleaching, fluorescent protein maturation and, crucially, protein degradation.

The first finding is a matter of minutes. Upon removal of the carbon source, every cell stopped growing essentially immediately, within the three-minute resolution of the microscopy. Contrary to any picture of the stationary phase as a mixture of growing and dying cells, the population entered growth arrest as one. Most remarkably, after growth had already ceased, roughly 38 percent of cells went on to divide during the following ten hours, undergoing what biologists call reductive division. These divisions shrank the average cell size of the population and occurred preferentially in longer cells, likely those that had already progressed furthest in their cell cycle, perhaps having started or completed DNA replication. Even after 60 hours without food, more than 90 percent of cells were still capable of regrowing when nutrients returned.

With growth arrested, attention turned to gene expression. Here the diversity across promoters was dramatic. Ribosomal protein promoters such as rplN, which drive the cell’s protein synthesis machinery during exponential growth, were shut down almost completely within an hour of the switch. Other promoters, like hupA, showed a sharp initial drop followed by a slower, sustained trickle of production. A third class, exemplified by hslV and the RpoS-regulated promoter bolA, did something counterintuitive: production briefly increased, sometimes above its level during growth, before decaying exponentially over the following hours. By late starvation, the relative production rates across the 22 promoters spanned more than two orders of magnitude, from about 20 percent down to 0.1 percent of their growth-phase values, underscoring how profoundly the regulatory network is rewired when nutrients vanish.

What makes these dynamics especially noteworthy is their consistency across cells. The researchers had anticipated that individual bacteria might respond to starvation in idiosyncratic ways, as has been observed for other stresses such as the SOS response to DNA damage, where the population-level induction conceals almost no coordination between cells. Instead, during early starvation, single cells followed nearly identical trajectories. For promoters showing a burst of expression, more than half of all the variance observed across cells and time was captured simply by the population average, a degree of coordination rarely seen in bacterial gene regulation. This homogeneity suggests that entry into starvation triggers a tightly controlled program rather than a stochastic drift into dormancy.

Production rates, however, are only half the story; protein concentrations are what determine physiology. Because growth arrest halts the dilution of existing proteins while degradation continues, and because production does not stop instantly, many proteins accumulated substantially even as their synthesis slowed. The researchers measured the degradation rate of fluorescent proteins directly and found that it too decays exponentially during starvation, dropping from initial values around 0.02 to 0.03 per hour. This slowdown makes mechanistic sense: most protein degradation in E. coli is carried out by ATP-dependent proteases, and ATP levels plummet when carbon is gone. Integrating production, dilution and degradation in a simple mathematical model, the team showed that promoter concentrations either fell by as much as half or rose by up to twelvefold, with nearly all of these changes occurring within the first five to ten hours. After that, protein levels were essentially frozen, locking in a phenotype that persisted for days.

The model also revealed an elegant set of principles governing the final proteome. The concentration fold-change for each gene depends on just three effective ratios: the total degradation capacity, the balance between the fold-change in production and the fold-change in degradation relative to growth rate, and the relative time scales over which production and degradation decay. Even without a burst of expression, a promoter whose production decays gradually over about five hours can achieve a fourfold increase in protein concentration, simply because dilution has stopped while synthesis continues. For RpoS-regulated promoters such as bolA, an initial burst pushes the increase further, to around sevenfold. Experiments with an rpoS deletion strain confirmed the central role of this alternative sigma factor, although not all starvation-induced promoters depended on it, pointing to additional regulators still to be identified.

Importantly, the response did not require cells to anticipate starvation in advance. In a separate set of experiments, the researchers let a batch culture exhaust its glucose gradually while flowing the same medium through the microfluidic chip, so that cells inside the chip experienced a slower, roughly 45-minute transition into growth arrest instead of an abrupt five-minute switch. The gene expression dynamics and the resulting protein concentration changes were remarkably similar between the gradual and abrupt transitions, suggesting that the program is triggered by the cessation of growth itself rather than by sensing dwindling nutrients. Even when protein synthesis was temporarily blocked with chloramphenicol during the first five hours of starvation, the expression program simply resumed once the drug was removed, implying that the eventual decline in production reflects exhaustion of internal resources rather than a specific transcriptional shutoff.

The most consequential result concerns what this early remodeling is for. When the researchers subjected starving cells to a strong oxidative stress with hydrogen peroxide at the 20-hour mark, nearly 98 percent of normally starved cells survived. But when translation was inhibited from the very start of starvation, survival fell to 19 percent, whereas inhibiting protein production only after the first five hours was far less damaging, allowing 65 percent of cells to survive. The protective early expression tracked the accumulation of RpoS-regulated reporters, and a strain lacking rpoS altogether fared even worse than translation-inhibited wild-type cells, with survival dropping almost to zero. Together, these experiments demonstrate that the protein production of the first five hours is what arms the cells against stresses arriving much later, when their capacity to respond has essentially vanished.

The findings also suggest how starving cells manage to keep producing proteins at all. With ribosomal gene expression shut off, the reduced translation capacity is redirected away from growth machinery toward protective proteins, effectively raising the relative output of many genes even as total synthesis falls. The energy likely comes from endogenous metabolism, the controlled self-digestion of abundant cellular constituents: the team’s degradation measurements imply that roughly 40 percent of the proteome may be broken down within the first day of starvation, releasing metabolites and fueling ATP production. In this light, the first hours of starvation emerge as a resource-limited but precisely coordinated investment phase, in which a bacterium cannibalizes part of itself to build the molecular toolkit it will need to endure an uncertain and potentially hostile future. For a organism long caricatured as a simple growth machine, E. coli turns out to be a far more deliberate survivor than its reputation suggests.

Subject of Research: Proteome remodeling and gene expression dynamics in single E. coli cells during carbon starvation

Article Title: E. coli prepares for starvation by dramatically remodeling its proteome in the first hours after loss of nutrients

Article References: E. coli prepares for starvation by dramatically remodeling its proteome in the first hours after loss of nutrients. (n.d.). https://doi.org/10.1038/s44320-026-00226-5

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00226-5

Keywords: E. coli, carbon starvation, proteome remodeling, gene expression, microfluidics, single-cell microscopy, RpoS, stationary phase, protein degradation, stress tolerance, bacterial survival, Molecular Systems Biology

The researchers set out to close a major gap in kinome pharmacology. Protein kinases—the 518 enzymes that phosphorylate most human proteins and orchestrate nearly every cellular process—are not merely catalytic cores. Beyond their conserved kinase domains, they carry additional domains that mediate autoinhibition, subcellular localization, and complex formation. Autoinhibitory domains (AIDs) typically keep kinases dormant by docking onto the kinase domain and masking the ATP-binding site, blocking both enzymatic activity and substrate interactions until an activating signal relieves this restraint. Classical structural methods struggle to characterize these domains because many AIDs are intrinsically disordered or connected to the catalytic core through flexible linkers, and most structural studies rely on truncated, purified recombinant proteins that lack the physiological post-translational modifications and binding partners essential for correct function. Whether ATP-competitive inhibitors, which stabilize the active DFG-in conformation of the kinase domain, also force structural changes at the AID remained a poorly explored dark space.

To address this, the team developed a multimodal proteomics strategy combining three complementary mass spectrometry-based approaches. The first, AP-LiP-MS, applies limited proteolysis coupled to mass spectrometry on affinity-purified samples: kinases are purified from human cells under native conditions, treated with an inhibitor, and then exposed to proteinase K, which preferentially cleaves accessible and flexible regions. Changes in the resulting conformation-specific peptide fragments reveal structural shifts with high sequence coverage while preserving the native cellular context. This structural readout was paired with contextual proteomics—affinity purification mass spectrometry (AP-MS) to measure complex formation and in vivo proximity labeling using the miniTurbo biotin ligase to map the kinase’s biochemical neighborhood in living cells. The approach was first benchmarked on the kinase DCLK1, where known X-ray crystal structures of the autoinhibited and inhibitor-bound states confirmed that AP-LiP-MS faithfully detects the anticipated structural rearrangements.

Applying the workflow to three disease-associated kinases with well-characterized ATP-competitive inhibitors—CAMKK2 targeted by SGC-CAMKK2-1, CHEK1 targeted by rabusertib, and PRKCA targeted by Gö 6983—the researchers found a striking common theme. In every case, inhibitor binding produced structural changes precisely at the autoinhibitory domain, with increased proteinase K susceptibility indicating that the AID becomes more solvent-exposed. This is consistent with the AID dissociating from the kinase domain, driving the inhibited enzyme into an open, active-like conformation that mimics the structural unlocking that occurs during normal kinase activation. Notably, a structurally similar negative control compound that does not bind CAMKK2 induced neither structural nor interaction changes, confirming the specificity of the observations. For PRKCA, structural alterations extended into the membrane-binding C2 domain, specifically at a short regulatory segment associated with autoinhibition, hinting that multiple domain-domain interactions are disrupted by drug binding.

The consequences of these conformational shifts proved to be as diverse as they were unexpected. For CAMKK2, the inhibitor stabilized a complex between CAMKK2 and PRKAA1, the catalytic subunit of the AMPK energy-sensing complex. Catalytically inactive and autonomously active CAMKK2 mutants responded to the drug with the same interaction pattern, demonstrating that the effect depends on the conformational change rather than on catalytic inhibition. Structural modeling with AlphaFold3 suggested that the activation-relevant T183 residue of PRKAA1 becomes buried within the predicted CAMKK2–PRKAA1 interface. Functional experiments confirmed the implication: in glucose-starved cells, SGC-CAMKK2-1 reduced T183 phosphorylation of PRKAA1 by upstream kinases such as LKB1, and this suppression was rescued when CAMKK2 was depleted by siRNA. In other words, the inhibited kinase acts as a physical shield that sequesters AMPK and blocks its activation through an entirely non-catalytic, scaffolding mechanism—potentially shutting down both the calcium-dependent and energy-stress branches of AMPK signaling simultaneously.

Strikingly, a disease-associated CAMKK2 variant, the R311C mutation found in a patient with bipolar disorder, completely abolished the inhibitor-induced interaction with PRKAA1. Because R311 faces the predicted interaction interface while the neighboring catalytic residue D312 lies outside it, the finding offers the first mechanistic clue for how this genetic variant may uncouple the CAMKK2–AMPK signaling axis in patients, and it underscores that drug responses can depend critically on the specific disease variant a patient carries—a consideration for personalized medicine.

The second model kinase, CHEK1, revealed a different flavor of paradox. Rabusertib remodeled CHEK1’s interactions with numerous DNA-damage response proteins, increasing binding to 14-3-3 proteins, the deubiquitylating enzyme USP7, PCNA, and MCM replication licensing factors, and elevating phosphorylation at the ATR-targeted S317 site—changes mirroring those seen during genuine DNA damage-induced activation. The single interactor that dissociated was CLPB, a mitochondrial protein previously identified in multiple studies as a CHEK1 partner. CLPB dissociation occurred in both wild-type and catalytically inactive CHEK1 but not in a constitutively open mutant, again implicating the conformational rather than catalytic consequence of inhibition. Because CLPB loss is known to cause mitochondrial fragmentation, the researchers examined mitochondrial morphology by super-resolution microscopy. Rabusertib treatment significantly increased mitochondrial fragmentation, an effect that persisted even when the canonical CDK1–DRP1 fragmentation pathway was blocked with the CDK1 inhibitor RO-3306, and that could not be reproduced by DNA damage alone. While a direct causal link between CHEK1–CLPB dissociation and fragmentation remains to be established, the data suggest that CHEK1 inhibition may disrupt mitochondrial proteostasis through a mechanism independent of the drug’s intended catalytic target.

The third model, PRKCA, demonstrated how inhibitor-induced structural changes can redirect a kinase within the cell. Upon Gö 6983 binding, proximity labeling revealed a rapid shift of PRKCA toward membrane-associated proteins at cell junctions, including tight junction, adherens junction, and desmosome components, as well as the known interactor integrin beta-1. Calcium imaging ruled out changes in intracellular calcium as the driver, and a dose-response experiment showed that these junctional proximity changes occurred at significantly lower drug concentrations than other effects, consistent with a specific on-target mechanism. Catalytically inactive and constitutively active PRKCA mutants responded identically, confirming the phenotype is independent of catalytic inhibition. Live-cell imaging of EGFP-tagged PRKCA captured the kinase relocating to the cell periphery—particularly cell-cell contact sites—within eight minutes of drug addition. The researchers propose that Gö 6983 binding opens the C2 domain, exposing a lysine cluster that can bind the junctional lipid PIP2, thereby recruiting the inhibited kinase to membranes through a calcium-independent route.

Taken together, the study establishes the ATP-binding site as a major organizing center of kinase conformation and interaction, and suggests that inhibitor-induced non-catalytic gain-of-function is likely far more prevalent among kinases with autoinhibitory domains than currently appreciated. The authors point to existing examples such as the JAK2 inhibitor ruxolitinib, which paradoxically primes JAK2 hyperphosphorylation and contributes to side effects after drug withdrawal, and BRAF inhibitors that allosterically promote RAF dimerization and MAPK activation, as evidence that such mechanisms already matter clinically. Because many AID-mediated effects would be missed by conventional target engagement assays focused on catalytic kinetics, the authors advocate for systematic multimodal proteomic profiling of both wild-type and disease-mutant kinases during early drug development. The combination of structural and contextual proteomics is not restricted to kinases and could extend to targets lacking catalytic activity altogether. By mapping inhibitor-modulated conformational and interactome landscapes early, researchers hope to detect unexpected liabilities before they surface in the clinic, ultimately guiding the design of safer and more effective kinase-targeted therapeutics.

Subject of Research: Inhibitor-induced displacement of kinase autoinhibitory domains driving non-catalytic drug effects

Article Title: Paradoxical non-catalytic kinase functions are driven by inhibitor-induced displacement of autoinhibitory domains

Article References: Reber, V., Keller, S., Loosli, S. A., Arima, Y., Kleele, T., Picotti, P., & Gstaiger, M. (2026). Paradoxical non-catalytic kinase functions are driven by inhibitor-induced displacement of autoinhibitory domains. Molecular Systems Biology, 22(9), 1474-1500. https://doi.org/10.1038/s44320-026-00229-2

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00229-2

Keywords: kinase inhibitors, autoinhibitory domains, proteomics, protein-protein interactions, CAMKK2, CHEK1, PRKCA, AMPK, mitochondrial fragmentation, drug mechanisms, conformational change, Molecular Systems Biology

Swarming is a collective form of surface motility in which bacteria elongate, hyperflagellate and coordinate their movement to spread rapidly across surfaces. As a colony expands, its morphology integrates cellular responses over space and time, effectively producing a visible record of the conditions the cells encountered during growth. The team reasoned that this emergent behavior could be hijacked for information recording, complementing DNA-based recorders that store data in genomic sequences and require sequencing to read out. Unlike molecular recorders, swarm patterns can be inspected with the naked eye or a cheap flatbed scanner.

The platform is anchored in a hypermotile derivative of the canonical E. coli K-12 strain MG1655, known as MG1655hm, which carries an insertion sequence upstream of the flhD promoter that boosts expression of the master flagellar regulator flhDC. On semi-solid agar supplemented with glucose, this strain forms reproducible, radially symmetric colonies spanning several centimeters within 24 hours. To make the patterns input-responsive, the researchers built high-copy plasmids in which an IPTG-inducible pLac promoter drives the expression of individual swarming-related genes, with green fluorescent protein serving as a negative control.

Four candidate genes were selected for their distinct positions in the motility regulatory hierarchy. The leucine-responsive regulatory protein lrp sits near the top, activating flhDC; fliA encodes the flagellum-specific sigma factor sigma-28 that governs late flagellar genes; rpoS encodes the stress sigma factor sigma-S, which indirectly represses flagellar expression; and cheZ encodes the chemotaxis phosphatase that terminates rotational signaling and promotes forward runs. Overexpression of lrp produced feather-like branching and larger colonies, fliA yielded smaller, denser colonies with jagged borders, and rpoS generated compact colonies with a dense inner ring and short protrusions. CheZ overexpression, by contrast, left colony morphology essentially unchanged.

When the team swept IPTG concentrations across 24-hour swarming assays, the strains fell into two functional classes. The pLac-lrp strain behaved like a threshold-based, binary recorder: below roughly 0.1 millimolar IPTG it produced the baseline colony, while above that concentration its swarming expanse jumped dramatically to cover the entire Petri dish, with step-like changes in density variation and low-frequency image features. The pLac-fliA and pLac-rpoS strains, in contrast, responded in a graded, analog-like fashion, with colony radius, texture, eccentricity and pixel intensity shifting progressively across wide inducer ranges. Control strains carrying only gfp showed no meaningful pattern changes.

To read the recorded information, the researchers developed a semi-automated image-analysis pipeline that segments colony scans, converts them into Cartesian, polar and Fourier representations, and extracts features of scale, shape, density and texture. Regression models fit to these features decoded IPTG levels with striking fidelity: the binary lrp model achieved 100 percent accuracy with an area under the ROC curve of 1, while the analog fliA and rpoS models reached five-class decoding accuracies of 94 and 78 percent, with AUCs of 1 and 0.96. Control models performed near chance, confirming that decoding success depended on genuine engineered responses.

The team also trained deep-learning classifiers on raw colony images using the SwinTransformer vision architecture, bypassing handcrafted feature extraction entirely. These models matched or approached the regression results, with the lrp model again achieving perfect binary classification and the fliA and rpoS models reaching accuracies of 92 and 82 percent. Most notably, a single unified model trained on pooled images from both analog strains simultaneously decoded strain identity and IPTG concentration across nine classes, achieving 93 percent accuracy and a macro-averaged AUC of 0.9986, demonstrating that multiplexed sensing does not come at the cost of decoding performance.

Because swarming is inherently dynamic, the researchers repurposed a commercial flatbed scanner placed inside a 37-degree incubator to capture images every ten minutes throughout colony development. Time-lapse analysis revealed that expansion rate and local textural uniformity recorded IPTG dose in real time, and statistical power analyses showed that the two analog strains could be distinguished from uninduced controls with near-total confidence as early as five hours into growth. The platform could even register transient events: when a single drop of concentrated IPTG was pipetted beside an expanding colony front at 12 hours, rpoS colonies developed a measurable asymmetry, with reduced expansion specifically on the inducer-facing side, while control colonies remained radially symmetric.

To demonstrate modularity beyond chemical induction, the team swapped the IPTG circuit for pDawn, a blue-light optogenetic system, placing fliA under the control of a light-activated promoter. Illuminated colonies became progressively smaller and denser with increasing blue-light irradiance, and the same image-analysis pipeline, applied without modification, quantified graded analog responses in minimum radius and density heterogeneity. The authors argue that this plug-and-play design, combined with E. coli’s vast genetic toolkit, could extend swarm-based recording to environmental monitoring, clinical biosensing, educational demonstrations and engineered living materials, turning one of biology’s oldest collective behaviors into a programmable, macroscale readout of the microbial world.

Subject of Research: Engineering E. coli swarming motility to spatially record chemical and optical inputs in decodable colony patterns

Article Title: Engineered E. coli swarming for binary and analog input recording

Article References: Shaw, M., Garg, S., Kirby, J., Schmidgall, S., Liguori, F., Doshi, A., Tu, T., & Danino, T. (2026). Engineered E. coli swarming for binary and analog input recording. Molecular Systems Biology, 22(9), 1377-1400. https://doi.org/10.1038/s44320-026-00232-7

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00232-7

Keywords: E. coli, swarming motility, synthetic biology, biosensors, optogenetics, pattern formation, machine learning, image analysis, information recording, bacterial colonies, genetic circuits, Molecular Systems Biology