The core idea is a two-phase relay built on Talbi’s High-level Relay Hybridization taxonomy. In the first phase, one of three exploration-oriented metaheuristics searches the binary feature space: Particle Swarm Optimization (PSO), which mimics flocking behavior; Harris Hawks Optimization (HHO), which models cooperative hunting with an escaping-energy parameter that shifts the search from exploration to exploitation; or Brain Storm Optimization (BSO), which clusters candidate solutions the way human brainstormers group ideas. In the second phase, the Binary Al-Biruni Earth Radius algorithm (bABER) takes over. Inspired by the medieval scholar’s geometric method for estimating the Earth’s radius, bABER generates adaptive step sizes and splits its population into exploration and exploitation groups whose proportions shift linearly from 70 percent to 30 percent over its run, with a mutation operator rescuing any individual that stagnates for three consecutive iterations.

The link between the phases is a direct population-transfer strategy. Rather than reinitializing randomly or passing only summary information, the framework moves the entire evolved population from phase one into bABER, preserving individual-level structure and the search history embedded in it. Because the transfer copies every solution, it cannot by itself reduce population diversity; the real risk is stagnation during refinement, which bABER’s shifting exploration share and mutation mechanism are designed to counter. The transition point was set quantitatively rather than by eye: the base algorithms realize roughly 98 to 99 percent of their total fitness gain by iteration 60, so a 60/40 split of the 100-iteration budget leaves exploration essentially complete before refinement begins.

Every candidate subset is scored by a fitness function that weights classification accuracy at 0.99 and penalizes subset size with the remaining weight, evaluated by 5-fold cross-validation on the training partition with a fixed Random Forest classifier. The authors are unusually transparent about what that weighting actually buys. They derive an exchange rate showing how many features must be removed to offset a single cross-validation error: about 1.5 features per error on the Diabetes dataset, where the compactness term exerts genuine pressure, but over 16 on Parkinson’s, where it functions only as a tie-breaker among equally accurate subsets. Since all seven algorithms share the same objective, differences in subset size must come from search dynamics, not from the penalty itself.

Evaluation covered six benchmark medical datasets, from Parkinson’s disease voice measurements and PIMA diabetes records to breast cancer, cervical cancer, heart disease, and a 2,149-case Alzheimer’s dataset. The protocol is deliberately leakage-aware: 15 independent runs per algorithm per dataset, each with its own stratified 70/30 split, z-score scaling fitted only on training data, and no selection among runs using test performance. That yields 630 optimization runs in total, with all metrics reported as means and standard deviations. The authors even flag an implementation inconsistency they discovered: in PSO-bABER the phase-two schedule continues on the global iteration counter rather than resetting, an offset they report as observed rather than quietly re-running.

The statistical verdict is refreshingly sober. A Friedman test finds no significant difference in accuracy among the seven methods, and equivalence testing with a one-percentage-point margin confirms that each hybrid matches its base algorithm. The single nominally significant accuracy result, BSO-bABER over BSO at p = 0.031, carries a negligible effect size and does not survive Holm correction for multiple comparisons. What the statistics do support is an effect on compactness: the Friedman test on the number of selected features is significant at p = 0.015, and HHO-bABER selects fewer features than HHO on five of six datasets, with a mean reduction of about two features and the largest effect size in the study outside execution-time comparisons.

The clinical significance of that compactness effect should not be understated. HHO-bABER selects between roughly one and seven fewer features than HHO depending on the dataset, at unchanged accuracy, which translates directly into fewer measurements to acquire per patient when features are laboratory assays or biomedical voice recordings. On the imbalanced cervical cancer dataset, where a majority-class classifier would already score 93.4 percent accuracy, the hybrid variants also produce far more consistent subset sizes across runs than standalone bABER, whose aggressive reduction yields subsets varying by more than 100 percent of their mean. Reproducibility of the selected feature set matters for clinical trust, and the hybrids deliver it where the standalone algorithm does not.

There is a price. All three hybrids are significantly slower than their base algorithms on every dataset, with large effect sizes, because the refinement stage adds fitness evaluations dominated by cross-validated Random Forest training. HHO-bABER is the most expensive at roughly 881 seconds on average. Yet relative to standalone bABER, the PSO- and BSO-based hybrids are about 22 percent faster, and execution time scales mainly with the number of instances rather than features: an elevenfold increase in sample size across the benchmarks multiplied runtime by only about 1.3. A sensitivity analysis of the transition ratio found accuracy varies by at most about one percentage point across 40/60, 60/40, and 70/30 configurations, so the authors present their default as a reasonable choice within a broad plateau rather than an optimum.

The study is equally candid about its limits. The datasets span only 8 to 35 features, and the authors show that their scalarized fitness function would collapse to accuracy alone on high-dimensional gene-expression data, where the compactness penalty could never offset a single cross-validation error; a Pareto formulation or a filter stage would be needed there. Population states were not logged per iteration, so diversity preservation across the transfer could not be measured directly, and subset stability indices could not be computed retrospectively. All conclusions are conditional on a single fixed Random Forest, and validation on prospectively collected institutional clinical data remains necessary before any clinical claim.

What emerges is a template for how hybrid optimization studies should be reported. The framework itself is algorithm-agnostic: because the transfer operates on binary solution vectors rather than algorithm-specific internal states, any population-based explorer, from Grey Wolf to Whale Optimization, could be slotted into phase one without modification. The complete code, including split seeds and analysis scripts, is publicly available on GitHub and archived on Zenodo. In a field crowded with inflated claims of accuracy gains from algorithm stacking, this team’s conclusion that sequential hybridization is best understood as a way of obtaining more compact feature subsets at preserved accuracy, paid for in runtime, is exactly the kind of precise, falsifiable finding that makes medical machine learning incrementally more trustworthy.

Subject of Research: Sequential hybrid metaheuristic feature selection for medical machine learning using population transfer

Article Title: A sequential hybrid framework for medical feature selection using population transfer

Article References: Lazrak, C., Bouayad, A., & Talha, A. (2026). A sequential hybrid framework for medical feature selection using population transfer. Discover Artificial Intelligence, 6(1), Article 1429. https://doi.org/10.1007/s44163-026-02217-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02217-9

Keywords: feature selection, metaheuristics, particle swarm optimization, Harris Hawks optimization, brain storm optimization, Al-Biruni Earth Radius algorithm, medical machine learning, Random Forest, hybridization, data leakage, statistical validation, clinical diagnostics

The result is notable less for the raw accuracy figure, which in machine learning benchmarks can sometimes be inflated by favorable data splits, than for the trade-off the model appears to strike. Deep learning systems that read medical images have historically faced a frustrating compromise: the largest, most accurate models are often too slow and computationally hungry for real-time use, while smaller, faster models sacrifice diagnostic precision. The authors argue that DeiT-III threads this needle, outperforming a battery of competing architectures, including DeiT-Tiny, DeiT-Small, BeiT, ViT-B16, CrossViT-Tiny, and PiT-Ti, on both accuracy and inference time. That combination of high precision, low latency, and computational efficiency, they conclude, makes the framework a strong candidate for integration into real-time clinical pathology workflows.

To understand why this matters, it helps to consider what a vision transformer actually does differently from the convolutional neural networks that dominated medical image analysis for the past decade. Convolutional networks process images through layers of filters that scan for local patterns, edges, textures, and progressively more abstract shapes, building up a representation of the image neighborhood by neighborhood. Transformers, by contrast, borrow the self-attention mechanism that revolutionized natural language processing. An image is chopped into small patches, each patch is treated roughly like a word in a sentence, and the model learns which patches are relevant to which others, no matter how far apart they sit in the image. For histopathology, where a diagnostic clue in one corner of a tissue patch may only make sense in light of cellular architecture on the opposite side, this global attention mechanism is an appealing fit.

The catch with transformers has always been data hunger. The original Vision Transformer, introduced by Dosovitskiy and colleagues in 2020, famously underperformed comparable convolutional networks unless it was pre-trained on enormous datasets, on the order of hundreds of millions of labeled images. That is an unrealistic requirement in medicine, where annotated images are expensive, privacy constraints are severe, and expert pathologists’ time is scarce. The data-efficient image transformer family, including the DeiT-III variant used in this study, was engineered specifically to relax that requirement, using improved pre-training and regularization strategies so that the models can learn useful visual representations from far more modest corpora and then transfer that knowledge to new tasks.

That is precisely the strategy the research team employed. They initialized their model with transfer learning from ImageNet, the large general-purpose image dataset that has served as the standard starting point for computer vision research for years, and then fine-tuned the network on the LC25000 dataset, a publicly available collection of 25,000 histopathological images covering lung and colon tissue. For the lung cancer task, the relevant categories are benign lung tissue, lung adenocarcinoma, and lung squamous cell carcinoma, the two malignant subtypes that together account for the majority of non-small-cell lung cancers. Distinguishing among these three classes on stained tissue sections is a task that trained pathologists perform daily, and one where the stakes of an error, a missed malignancy or an unnecessary aggressive workup, are genuinely clinical.

The reported numbers are striking. A mean accuracy of 98.64 percent across the three classes means the model misclassified fewer than two images in a hundred, and an AUC of 0.9999 indicates that across essentially all decision thresholds, the model’s ranking of diseased versus non-diseased cases was nearly flawless on the test data. Just as important for practical deployment, the authors report that DeiT-III beat its competitors not only on accuracy but on inference time, the latency between feeding an image into the network and receiving a classification. In a pathology workflow where a single whole-slide image can contain billions of pixels and a lab may process thousands of slides a day, shaving milliseconds off per-patch inference compounds into hours of saved computation.

The study arrives amid a broader surge of interest in applying transformers to medical imaging. Recent reviews have catalogued vision transformer applications across diabetic retinopathy screening, tumor classification on MRI and CT, oral epithelial dysplasia grading, breast cancer histopathology, and lung nodule segmentation, often in hybrid architectures that pair a transformer’s global attention with a convolutional backbone’s local feature extraction. The field’s enthusiasm is easy to understand: histopathology images are enormous, structurally rich, and full of long-range spatial relationships that fixed convolutional kernels struggle to capture. But the literature also shows that no single architecture wins everywhere, which is why comparative studies like this one, pitting DeiT-III against BeiT, CrossViT, PiT, and vanilla ViT variants on the same data, are useful for practitioners deciding what to actually deploy.

Caveats remain, and the authors and the wider field are careful about them. The LC25000 dataset, while public and widely used, consists of image patches curated from whole-slide images, meaning the model was evaluated on relatively uniform, well-prepared tiles rather than the messy, artifact-laden, variably stained slides that characterize real diagnostic work. Performance on a curated benchmark does not automatically translate to performance across scanning devices, staining protocols, and patient populations. The authors also note that the code and pre-trained models are currently held in a private repository, with public release planned via the first author’s GitHub page, an important step for independent validation. Regulatory approval, prospective clinical testing, and integration with laboratory information systems all lie between a 98.64 percent benchmark score and a system a hospital can trust.

Interpretability is another open question the paper itself flags. The authors write that existing deep learning models usually involve a trade-off between computational efficiency and sound clinical interpretability, and while their framework addresses the efficiency side of that equation, convincing pathologists to trust a black-box classifier will require attention maps, uncertainty estimates, and validation studies that show the model is looking at the same cellular features a human expert would. The field of explainable AI in histopathology is advancing quickly, with dedicated transformer models featuring built-in explainability now appearing in the literature, and future work will need to bring those threads together.

Still, the trajectory is hard to ignore. In less than a decade, computational pathology has moved from early convolutional networks that could distinguish tumor from non-tumor, such as the landmark 2018 study by Coudray and colleagues that classified non-small-cell lung cancer subtypes and even predicted mutations from histology, to transformer models approaching the theoretical ceiling of three-class classification accuracy. If data-efficient architectures like DeiT-III can hold their performance outside the benchmark and deliver their speed advantage in production hardware, the vision of an AI assistant sitting beside the pathologist, flagging suspicious patches in real time and never tiring at the end of a long day of slide reading, moves considerably closer to reality. For a disease that kills more people than any other cancer, that is a development worth watching closely.

Subject of Research: Data-efficient vision transformer classification of lung cancer histopathology images

Article Title: Data-efficient image transformer (DeiT) for high accuracy lung cancer histopathology classification

Article References: Hussain, S., Wenhui, Y., Hussain, B., Hussain, E., & Jin, H. (2026). Data-efficient image transformer (DeiT) for high accuracy lung cancer histopathology classification. Multimedia Tools and Applications, 85(10), Article 798. https://doi.org/10.1007/s11042-026-21954-8

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21954-8

Keywords: lung cancer, histopathology, vision transformer, DeiT-III, deep learning, transfer learning, LC25000 dataset, medical imaging, computational pathology, machine learning, image classification, clinical diagnostics

The starting point for the framework is a sobering observation from the change management literature: intentional change processes are frequently slow, meet resistance and fail outright, and they are a significant source of stress for employees. Scholars such as McFillen and colleagues have speculated that the failure to conduct an accurate organisational diagnosis, and the related failure to build readiness for change, plays a major role in the high rate of failed initiatives. Yet the diagnostic models most widely used in practice, including the McKinsey 7 S model, Weisbord’s six-box model and the Burke–Litwin model, have been criticised for presenting heterogeneous, atomistic categories without specifying how the factors interact, for omitting elements such as the external environment, and for lacking strong theoretical foundations. They offer managers step-by-step checklists, the authors argue, but little insight into why resistance arises or what lies beneath surface phenomena.

SyMOA is grounded in the systems psychodynamic tradition that emerged at the Tavistock Institute in London in the 1950s and 1960s. That approach combines open-systems theory, which sees organisations as self-sustaining systems in constant exchange with their environment, with Freudian and object-relations theories, which hold that unconscious processes, hidden conflicts and anxieties powerfully shape human behaviour. As researcher Yiannis Gabriel has succinctly put it, in a psychodynamic view conflict is unavoidable and constitutive of the subject. Systems psychodynamic scholars examine how organisations develop mechanisms to manage competing drives, such as the desire for change versus the desire for stability, or the need for independence versus the need for belonging, and they focus on group behaviour and the collective unconscious rather than isolated individuals. What has been missing, the Freiburg team contends, is a structured diagnostic system built directly from these insights.

To fill that gap, the authors turned to clinical practice. Their template is the Operationalised Psychodynamic Diagnosis, or OPD, a well-researched multiaxial system used in German-speaking psychotherapy that captures a patient’s psychological state across experience of illness, interpersonal relationships, life-determining unconscious conflicts and structural functioning. The OPD is typically administered through a semi-structured interview, and the resulting psychodynamic profile informs therapy planning and evaluation. The researchers stress that transferring psychodynamic concepts from the individual to the organisational level does not mean mapping intrapsychic processes onto companies. Instead, following the systems psychodynamic tradition, organisational structures, routines and relational patterns are understood as socially organised responses to anxiety, uncertainty and conflict, an abstraction of underlying principles rather than a literal transfer.

The resulting framework assesses organisations along three interrelated dimensions. The first, the organisation’s inner functioning, examines current challenges and how they are attributed, whether to technology, workforce, culture or an integrated socio-technical system, along with resources, barriers to change and organisational climate. Beneath these visible features lie the inner workings: regular, repetitive conscious and unconscious patterns through which the organisation deals with itself and the external world. These are rated across five categories, each considered both inwardly and toward the environment: internal and external awareness, regulatory capacity, communication, attachment, and organisational defence mechanisms. These workings change slowly over time and can be rated from well functioning to low functioning.

The second dimension addresses internal relational patterns and dynamics, which are shaped by organisational structure and the demands of communication, cooperation and coordination, but also by interpersonal affect and the history of shared interaction. Assessment focuses on observable relational behaviour and conscious emotional experience in specific situations, at horizontal levels such as between teams and vertical levels such as between senior management and employees. The approach explicitly captures how each party in an interaction experiences itself and the other, illuminating recurring and potentially dysfunctional patterns of relationship.

The third dimension probes the organisation’s unconscious conflicts. Just as people carry an inner experience shaped by early experiences and proven survival strategies, the authors argue, companies accumulate conflicts that begin as visible tensions between teams or managers and gradually shift inward, becoming part of the organisational DNA. Conflict here does not mean open dispute but the struggle between contradictory needs, such as how much freedom versus how much guidance, or how much responsibility versus how much security. When such basic conflicts are adequately resolved, an organisation can flex situationally between the poles of need; when it cannot, it rigidly pursues one pole, and the conflict is repeatedly activated by situational factors during the organisation’s development.

Conceptually, SyMOA is framed by a metamodern sensibility, drawing on the philosophy of Vermeulen and van den Akker, which holds modernist rigor and postmodern reflexivity in productive tension rather than resolving them. In practice this means organisational phenomena are analysed using structured dimensions while remaining open to multiple, potentially conflicting interpretations. The framework also aligns with critical theory’s project of revealing entrenched power relations: organisational culture, the authors note, can operate both as a collective social defence against anxiety, in the tradition of Menzies and Schein’s concept of learning anxiety, and as a subtle form of control in which employees internalise corporate values and police themselves. Through reification, historically contingent power relations become naturalised and appear as objective features of organisational life, resistant to critique. By surfacing hidden defences and covert emotional processes, SyMOA is intended to make these barriers to change visible and contestable.

The authors are careful to position the work as deliberately conceptual. SyMOA is not a universal or prescriptive model but one possible analytical framework, and its dimensions are conceptual categories derived from systems psychodynamic theory rather than validated measures. A companion study protocol published in JMIR Research Protocols outlines the next phases: a structured interview guide, application in instrumental case studies, and testing of intercoder reliability, construct validity, content validity and known-groups validity, with safeguards including informed consent, confidentiality, peer debriefing and triangulation to manage the ethical risks of deep, potentially sensitive inquiry. The authors also draw an analogy to psychotherapy research, where interventions succeed more often when structural aspects, relationship dynamics and conflicts are assessed in a differentiated way and matched to a client’s stage of readiness for change, from pre-contemplation through maintenance. If that logic holds for organisations, they suggest, a SyMOA profile could indicate an organisation’s readiness for change and help tailor interventions accordingly. Whether the framework can deliver on that promise will depend on the empirical studies now underway, but the ambition is clear: to give organisations a clinical-grade instrument for understanding not just what they do, but why they keep doing it.

Subject of Research: A systems psychodynamic diagnostic framework for assessing unconscious organisational dynamics

Article Title: A systems psychodynamic framework for understanding and assessing organisational dynamics

Article References: Lahmann, C., Rank, O., & Müller, A.-M. (2026). A systems psychodynamic framework for understanding and assessing organisational dynamics. Discover Psychology, 6(1), Article 272. https://doi.org/10.1007/s44202-026-00885-3

Image Credits: AI Generated

DOI: 10.1007/s44202-026-00885-3

Keywords: organisational diagnostics, systems psychodynamics, organisational change, psychodynamic theory, organisational culture, clinical diagnostics, metamodernism, critical theory, organisational psychology, change management, OPD, organisational dynamics

The inspiration comes directly from living cells. Biological ion channels are not rigid sieves; they are dynamic molecular machines that couple their local electrostatic environment to the hydration state of the interface, achieving selectivity and transport rates that approach the diffusion limit. Potassium channels, for example, use coordinated amino acid residues to strip the hydration shell from ions as they pass. Synthetic solid-state nanochannels have long tried to imitate this behavior by decorating rigid channel walls with stimuli-responsive molecules, converting molecular recognition events directly into measurable ionic currents. The problem has always been that these functionalizations are essentially two-dimensional, painting the sensing chemistry onto a flat surface where physiological salt can neutralize it.

The new platform, developed by Liu Shi, Genxi Li and colleagues, flips that geometry. Instead of modifying the channel surface, the team assembled a peptide-DNA hybrid hydrogel asymmetrically on the exterior of an anodic aluminum oxide membrane, creating a volumetric functional zone roughly 25 to 30 micrometers thick that acts as an active ion trap. The hydrogel is superhydrophilic and densely charged, so it pre-concentrates bulk ions through macroscopic Donnan enrichment while simultaneously generating a substantial Donnan potential. Because this charge reservoir occupies a volume orders of magnitude larger than the compressed Debye length, it keeps working even in the harshest ionic environments the body can supply.

The molecular architecture of the gate is a piece of careful supramolecular engineering. Single-stranded linker DNA is covalently anchored to the nanochannel surface, then a bridging strand and polyacrylamide conjugates are hybridized stepwise into the network. At the heart of the assembly sits a peptide-DNA complex cross-linker, synthesized through orthogonal thiol-maleimide coupling and strain-promoted alkyne-azide cycloaddition. The peptide backbone physically tethers the polymer-nucleic acid strands in close proximity, and this localized proximity effect entropically stabilizes the short DNA duplexes that hold the whole gel together. The elegance of the design is that the same tether that builds the network becomes its Achilles heel: a specific enzyme can cut it, and the entire structure falls apart on command.

That enzyme is Granzyme B, a serine protease released by cytotoxic T lymphocytes and a pivotal biomarker for gauging whether lung cancer patients are responding to immune checkpoint blockade therapy. When Granzyme B cleaves the peptide cross-linkers, the hydrogel undergoes a macroscopic gel-to-sol phase transition. Rheological measurements captured the collapse dramatically, with the storage modulus falling below the loss modulus as the elastic network vanished. Confocal microscopy showed fluorescently labeled gels fading after enzymatic treatment, while spectroscopy confirmed the loss of amide and hydrogen-bonding signatures. The gate does not merely weaken; it switches off.

The shutdown is synergistic, and that synergy is where the signal amplification comes from. As the gel disassembles, its volumetric charge reservoir dissipates, with the effective surface charge density dropping from -5.26 to -0.92 millivolts per square centimeter, an 82.5 percent loss in charge capacity. At the same time, the interface loses its superhydrophilicity, with the water contact angle rising from about 7.6 degrees to 39.0 degrees. Finite-element simulations governed by the Poisson-Nernst-Planck equations, fed with these experimentally measured parameters, showed the cation enrichment zone vanishing and the energy barrier for ion entry soaring. In a two-compartment electrochemical cell, the device swung from a high-current ON state of roughly 17.8 microamperes to an OFF state of about 3.07 microamperes, a gating ratio of approximately 5.8.

The comparison with a conventional two-dimensional control is what makes the result compelling. A monolayer of the same peptide chemistry on the channel surface showed only minor changes in wettability and charge after cleavage, and its sensing performance suffered accordingly, with a constrained dynamic range and reduced linearity. More tellingly, when the researchers varied the background ionic strength from 10 to 200 millimolar, the 2D device’s gating ratio collapsed sharply around 50 millimolar, exactly where Debye screening takes over. The 3D hydrogel platform, by contrast, maintained robust signal modulation all the way to 200 millimolar. In a physiological 150 millimolar buffer, the device still achieved a limit of detection of 133.60 femtomolar, while in low-salt conditions it reached an extraordinary 0.830 femtomolar across a range spanning six orders of magnitude.

The modularity of the building blocks opened a further, almost playful dimension: molecular computing. By incorporating cross-linkers responsive to both Granzyme B and matrix metalloproteinase 2 into a single hydrogel, the team built a cascaded logic circuit with two parallel INHIBIT sub-circuits feeding an OR gate. Each enzyme can independently trigger the phase disassembly and generate a current response, but a specific inhibitor for each enzyme vetoes the process. The device executed a full truth table of Boolean operations on biochemical inputs, supported by molecular docking simulations of the enzyme-inhibitor pairs. It is a small demonstration, but it points toward sensors that do not merely detect a molecule but compute a diagnosis from a panel of interacting biological signals.

Robustness in messy biological fluids proved equally impressive. The intact hydrogel doubles as a physical shield against non-specific fouling, and the sensor held its performance in 10 percent fetal bovine serum and 10 percent human serum, showing strict selectivity against a panel of interfering proteins. The device survived five regeneration cycles with consistent current switching, showed negligible degradation over 15 days of storage, and achieved batch-to-batch reproducibility with a relative standard deviation of just 1.62 percent. These are the unglamorous metrics that decide whether a laboratory curiosity can ever become a clinical instrument, and the platform passed them.

The clinical proof of concept is the most striking part of the story. Working with serum samples from healthy donors and lung cancer patients collected before and after immune checkpoint blockade therapy, under a protocol approved by the Medical Ethics Committee of Shanghai Pulmonary Hospital, the team measured Granzyme B directly in unpurified specimens. Healthy individuals averaged around 0.8 picomolar, while pre-treatment patients showed suppressed levels of roughly 0.3 to 0.5 picomolar, reflecting the immunosuppressive microenvironment of advanced malignancies. After immunotherapy, concentrations surged to 2.0 to 2.5 picomolar, signaling restored T-cell cytotoxicity. Benchmarked against a commercial ELISA, the sensor achieved a correlation coefficient of 0.962, a negligible mean bias of 0.0423 picomolar by Bland-Altman analysis, and an area under the ROC curve of 0.980 for distinguishing healthy from pre-treatment patients, with near-perfect discrimination of therapeutic response. The researchers acknowledge that translation will require packing the electrochemistry into automated microfluidic cassettes and expanding the logic gates to wider biomarker panels, but the core demonstration stands: by thinking in three dimensions, a dissolving gel has done what flat chemistry could not, carrying ultrasensitive nanochannel sensing out of the buffer and into the blood.

Subject of Research: A three-dimensional hydrogel soft-gating strategy for solid-state nanochannel biosensors that circumvents Debye screening to enable ultrasensitive detection of Granzyme B for immunotherapy monitoring.

Article Title: Hydrogel Phase Transition‐Driven Soft Gating Circumvents Debye Screening for Advanced Biosensing

Article References: Shi, L., Zhang, Z., Li, B., Gao, Y., Zhang, R., Mu, Z., Ni, J., Bo, B., & Li, G. (2026). Hydrogel Phase Transition‐Driven Soft Gating Circumvents Debye Screening for Advanced Biosensing. Advanced Science, Article e78049. https://doi.org/10.1002/advs.78049

Image Credits: AI Generated

DOI: 10.1002/advs.78049

Keywords: nanochannel biosensing, hydrogel, Debye screening, Donnan enrichment, Granzyme B, peptide-DNA hybrid, iontronics, immune checkpoint blockade, lung cancer, biocomputing, phase transition, clinical diagnostics

The study, led by teams at University Medicine Greifswald in Germany and Utrecht University in the Netherlands, exploited the TIMES cohort, an acronym for Tracking Individuals Monthly for Evaluating Stability. Participants, 32 women and 19 men with a mean age of about 34 years, donated blood once a month for twelve months between June 2017 and May 2018. This design is unusual in plasma proteomics, where most biomarker studies still rely on cross-sectional snapshots of a single time point per subject. By sampling repeatedly, the researchers could separate the variation that occurs within a person over time from the variation that exists between people, a distinction that is fundamental to interpreting any clinical measurement.

What makes the study particularly persuasive is its dual-laboratory architecture. The Greifswald site and the Utrecht site processed and analyzed the same samples using deliberately different workflows. One laboratory used a bead-based SP3 digestion protocol, the other a conventional liquid-phase digest. One measured peptides on a timsTOF Pro2 mass spectrometer, the other on the newer timsTOF HT, which features a fourth-generation trapped ion mobility analyzer with fivefold greater ion charge capacity. Even the software pipelines diverged: Greifswald searched raw data with Spectronaut against a large database of more than 42,000 human protein entries, while Utrecht used DIA-NN with a compact plasma-specific database of just 2,445 entries built from the Human Protein Atlas. Quantification strategies differed too, with Greifswald favoring intensity-based absolute quantification and Utrecht using MaxLFQ normalization.

Despite this deliberate methodological divergence, the two datasets agreed remarkably well. After filtering out contaminants and the hypervariable regions of immunoglobulins, Greifswald quantified 274 proteins and Utrecht 188, with 163 proteins overlapping. Comparing the two laboratories sample by sample, the mean Spearman correlation coefficient across all samples was 0.908 and the mean Pearson correlation was 0.824, indicating strong agreement in both the ranking of proteins and their absolute abundances. To make the numbers clinically meaningful, both sites converted relative intensities into absolute concentrations using a panel of 37 housekeeping plasma proteins with well-established reference values. The resulting estimates landed squarely within known clinical ranges, a benchmark that mass spectrometry workflows have historically struggled to meet.

With the analytical foundation secured, the team turned to the central biological question: how stable is a person’s plasma proteome over a year, and how different are people from one another? The answer came vividly from an unsupervised dimensionality reduction technique called UMAP, which projects high-dimensional protein data into two dimensions while preserving neighborhood structure. When all monthly samples from all donors were projected, each person’s twelve samples clustered tightly together, while the clusters of different donors sat far apart. In other words, your plasma proteome in January looks far more like your own proteome in December than like anyone else’s at any time of year.

To quantify this, the researchers built a machine learning classifier, a support vector machine with a radial basis kernel, and challenged it to identify which donor a blinded sample came from. Across 1,000 iterations, the model achieved a median classification accuracy of 98 percent. The proteome, in effect, carries a personal signature as distinctive as a fingerprint. Notably, statistical adjustment for age, sex and body mass index did not significantly reduce the inter-donor variability, suggesting that these common demographic confounders do not drive the observed differences in this cohort. The clustering was so robust that it even exposed a suspected sample mix-up: the December sample labeled as donor 27 sat squarely within donor 4’s cluster, prompting the team to exclude it as a probable aliquoting or labeling error.

The most consequential finding concerns the immunoglobulins, the antibody family that constitutes roughly 20 percent of total plasma protein mass. Because antibodies are extraordinarily polymorphic, with variable regions numbering in the millions of possible sequences, most plasma proteomics studies either physically deplete them or quietly ignore them in data analysis. This study did the opposite, retaining and quantifying them using peptides unique to the constant regions of each antibody class and subclass. The measurements showed that immunoglobulins vary enormously between donors but barely at all within a donor over twelve months. Total IgG averaged 9.75 grams per liter across the cohort with a standard deviation of 2.64 grams per liter, a 27 percent coefficient of variation, while the average intra-donor fluctuation was only about 1.11 grams per liter. The same pattern held for IgG subclasses IgG1 through IgG4, as well as IgM, IgA1, IgA2 and IgD, and it was mirrored by CD5L, a protein known to travel in complex with IgM.

The mass spectrometry-derived subclass distribution also matched the literature closely. IgG1 averaged 6.12 milligrams per liter of blood plasma, representing about 62.8 percent of total IgG, compared with a literature value of roughly 63.6 percent. IgG2 came in at 19.1 percent versus an expected 24.5 percent, IgG3 at 9.3 percent versus 5.7 percent, and IgG4 at 8.7 percent versus 6.3 percent. These numbers matter because they demonstrate that untargeted mass spectrometry can now deliver antibody measurements of clinical grade without the targeted immunoassays traditionally required. They also align with recent work showing that the dominant antibody clones in plasma are produced by long-lived plasma cells, which would explain why circulating immunoglobulin levels remain so stable over months.

Not everything in plasma stood still. C-reactive protein, the classic inflammation marker, behaved in a completely different manner. Most donors kept CRP concentrations low throughout the year, but a few individuals showed dramatic spikes, with one donor ranging from a basal 0.16 nanomolar to a peak of 721 nanomolar, a roughly 4,500-fold swing. Others, including six donors, maintained persistently elevated basal CRP of around 50 nanomolar across all twelve months, hinting at low-grade chronic inflammation. The episodic peaks were mirrored by sister inflammation markers SAA1, SAA2, S100A8 and S100A9, strengthening the interpretation that they reflect acute inflammatory episodes such as minor infections. The contrast between rock-steady antibodies and volatile CRP illustrates a key principle for biomarker research: some proteins define who you are, while others report what is happening to you right now.

The study also delivered an unplanned clinical vignette. One donor, identified as donor 26, showed immunoglobulin levels four to ten times below the cohort average for IgG1, IgG2, IgG3, IgA1 and IgM, all falling below the 2.5th percentile, combined with high body mass index and chronically elevated CRP. This pattern is suggestive of common variable immunodeficiency, the most prevalent primary immunodeficiency in humans, although the donor reported no increased susceptibility to infections. The researchers emphasize that this chance observation, surfaced purely through longitudinal proteomic screening, illustrates the diagnostic potential of the approach. More broadly, the work argues that biomarker discovery needs longitudinal designs, donor-specific baselines and cross-laboratory validation, and that a universal plasma reference material or isotopically labeled internal standards should accompany multi-center studies. If a single blood draw can be read against a person’s own stable proteomic baseline rather than a population average, the sensitivity of future diagnostics could improve substantially.

Subject of Research: Longitudinal mass spectrometry analysis of inter-donor variability in the human plasma proteome, focusing on immunoglobulin subclass stability

Article Title: Immunoglobulin sub-class levels define inter-donor plasma variability: a longitudinal dual-lab study

Article References: Michalik, S., Kalaidopoulou Nteak, S., Drouin, N., Gesell Salazar, M., Hammer, E., Dhople, V. M., Holtfreter, S., Weiss, S., van den Toorn, H. W. P., Bröker, B. M., Domańska, G., Völker, U., & Heck, A. J. R. (2026). Immunoglobulin sub-class levels define inter-donor plasma variability: a longitudinal dual-lab study. Molecular Systems Biology, 22(8), 1312-1332. https://doi.org/10.1038/s44320-026-00218-5

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00218-5

Keywords: plasma proteomics, immunoglobulins, mass spectrometry, longitudinal study, biomarkers, C-reactive protein, IgG subclasses, inter-donor variability, machine learning, clinical diagnostics, TIMES cohort, proteome stability

Mass spectrometry-based amyloid typing has earned a reputation as the gold standard for identifying the amyloidogenic protein in tissue deposits. By shredding laser-microdissected amyloid plaques into peptides and reading their mass spectra, laboratories can distinguish among more than thirty amyloid types with a specificity that immunohistochemistry and immunofluorescence simply cannot match. Yet the technique has remained confined largely to a handful of reference laboratories because it depends on nanoflow liquid chromatography, in which solvents are pushed through capillaries with inner diameters measured in tens of micrometers. Nanoflow systems deliver exquisite sensitivity, but they come with steep up-front costs, demanding maintenance, long column equilibration times, and a level of operational complexity that deters routine clinical adoption.

The Stanford team, led by Morgan W. Mann and Fangjun Chen with senior authors Megan L. Troxell and Ruben Y. Luo, set out to re-engineer the workflow so that it could run on the high-flow liquid chromatography systems already humming away in hospital laboratories. Their solution rests on two complementary innovations: a sample preparation method called suspension trapping, known commercially as S-Trap, and an acquisition strategy known as data-independent acquisition, or DIA. Together, these techniques allow the analysis of microdissected amyloid plaques using conventional-flow chromatography coupled to tandem mass spectrometry, without sacrificing the diagnostic accuracy that clinicians rely upon.

Suspension trapping is the quieter revolution of the two. In traditional proteomics sample preparation, proteins extracted from tiny tissue fragments must be digested into peptides through a series of solution exchanges, buffer removals, and cleanups that consume time and lose precious material at every step. The S-Trap approach instead traps proteins on a quartz filter matrix as they precipitate out of an acidic methanol-containing buffer. Digestion enzymes then pass through the trap, generating peptides that are eluted in a single, efficient step. The method is fast, robust, and minimizes sample loss, which matters enormously when the starting material consists of a few thousand cells harvested by laser capture microdissection from a formalin-fixed paraffin-embedded tissue section.

Data-independent acquisition addresses the other half of the problem. In classical data-dependent acquisition, the mass spectrometer selects the most intense ions in each scan for fragmentation, a stochastic process that can miss low-abundance peptides and produces data that are tedious to align across runs. DIA, by contrast, fragments all ions within successive wide mass windows, systematically recording fragment spectra for essentially everything in the sample. The resulting data are reconstructed computationally against spectral libraries, yielding consistent and quantitative peptide identifications. Because DIA tolerates the higher flow rates, wider chromatographic peaks, and greater sample loads of high-flow chromatography, it made the leap away from nanoflow feasible rather than merely aspirational.

To translate raw proteomic data into a diagnosis, the researchers also developed a custom selection heuristic that sifts through the hundreds of proteins detected in each plaque and identifies the most likely amyloidogenic culprit. Amyloid deposits are not pure; they sweep up serum proteins, immunoglobulin fragments, and extracellular matrix components as they form. Distinguishing the genuine amyloidogenic protein from innocent bystanders requires weighing total protein abundance, the presence of type-specific signature peptides, and the biological plausibility of each candidate. The heuristic automates this judgment, reducing an expert curation task that could take hours into a reproducible computational step suitable for a busy clinical workflow.

The performance data are striking. Working with laser capture microdissected plaques from 47 patient samples, 10 from cardiac biopsies and 37 from renal biopsies, the team split the material into a training set of 25 samples and an independent testing set of 22. The high-flow LC-DIA-MS/MS method identified the amyloidogenic protein in approximately 96 percent of plaques in both sets, a success rate comparable to established nanoflow approaches. The two inaccuracies observed across the cohort both traced to difficulties in identifying immunoglobulin lambda proteins, once in a specimen obtained after treatment had altered the composition of the deposits and once in a plaque containing two distinct amyloid types simultaneously. These edge cases, the authors note, define the boundaries of the method rather than undermining its general utility.

Reducing the sample preparation workflow to under eight hours carries practical significance beyond convenience. Amyloidosis is increasingly recognized as an urgent diagnosis; cardiac involvement can progress rapidly, and new therapies such as transthyretin stabilizers and silencing agents work best when started early. Every hour shaved from the analytical pipeline shortens the time between biopsy and answer. Equally important, the shift to high-flow chromatography means the method can run on instruments of the kind already deployed for clinical chemistry assays, drug monitoring, and newborn screening in hospital laboratories. That compatibility could transform amyloid typing from a rarefied referral test into an accessible diagnostic performed locally, at lower cost and with faster turnaround.

The study, published in Clinical Proteomics, was conducted on remnant patient specimens under approved institutional review board protocols and funded through the Stanford University Department of Pathology’s Test Development Program, underscoring its explicitly clinical orientation. The authors acknowledge one commercial entanglement: a collaborative research relationship between one investigator and Thermo Fisher Scientific, which provided technical support. While the method will still require validation in other laboratories and across broader panels of amyloid types, the combination of suspension trapping, data-independent acquisition, and high-flow chromatography represents a credible path toward democratizing molecular amyloid typing, offering patients everywhere the prospect of a faster, more precise answer to the question their treatment ultimately depends on.

Subject of Research: A high-flow liquid chromatography-tandem mass spectrometry method for typing amyloidosis in clinical laboratories

Article Title: Suspension trapping and data-independent acquisition enable high-flow liquid chromatography-tandem mass spectrometry-based amyloidosis typing in clinical laboratories

Article References: Mann, M. W., Chen, F., Zhu, C., Lu, C., Liang, B., Kambham, N., Troxell, M. L., & Luo, R. Y. (2026). Suspension trapping and data-independent acquisition enable high-flow liquid chromatography-tandem mass spectrometry-based amyloidosis typing in clinical laboratories. Clinical Proteomics. https://doi.org/10.1186/s12014-026-09628-x

Image Credits: AI Generated

DOI: 10.1186/s12014-026-09628-x

Keywords: amyloidosis, mass spectrometry, proteomics, data-independent acquisition, suspension trapping, liquid chromatography, clinical diagnostics, laser capture microdissection, AL amyloidosis, ATTR amyloidosis, tandem mass spectrometry, clinical proteomics