The investigation focused on GXPs, which entomopathogenic bacteria use to kill insect prey. The GxpS genes contain repeated domains whose separate histories are difficult to reconstruct because whole genes can undergo duplication or horizontal transfer, while individual domains and modules can also be duplicated, deleted, inserted or recombined. Conventional gene trees can therefore conflate several events. To establish a reference history, the researchers used the terminal thioesterase domain together with the adjacent T domain. Recombination-aware analyses detected no recombination in these regions among the sampled GxpS sequences, allowing them to serve as an anchor for the whole gene and for the individual A- and TC-domain trees.

Analysis of 54 putative GxpS orthologues placed them in a monophyletic group and included the four-modular Xenotetrapeptide synthetase and an unknown six-modular non-ribosomal peptide synthetase within that group. Reconciliation of the TTE-domain gene tree with a species tree indicated that GxpS first arose in Photorhabdus and was subsequently transferred to Xenorhabdus twice. This conflict between gene and species histories was interpreted as evidence of horizontal gene transfer. The researchers also traced the separate operon that supplies p-aminophenylalanine and its methylated derivative, concluding that it was present in the last common ancestor of the analyzed bacteria, was lost twice in Xenorhabdus groups, and was reacquired by horizontal transfer in Xenorhabdus doucetiae. In that species, the operon contains mobile genetic elements and experiments indicated that it was probably inactive.

Disagreement among the trees of the other GxpS domains exposed extensive recombination. Some exchanges replaced entire adenylation domains, whereas others transferred short sections, including fragments inserted into regions already altered by an earlier event. The team predicted 17 A-domain recombinations and used a mixture of phylogenetic incongruence, local sequence similarity and recombination-detection analysis to resolve more complicated histories. Their workflow separated unrecombined sequence from donor-derived segments and identified the likely donor clade for each inferred event. Experimental validation then included expression in Escherichia coli, high-performance liquid chromatography coupled to mass spectrometry, feeding with isotopically labelled amino acids, and co-injection of chemically synthesized peptides.

Those experiments confirmed every recombination that had been predicted to create a new peptide. They also established functional consequences that could not have been read directly from sequence alone. In one example, an A2-domain fragment replaced the A3 domain of the synthetase from Xenorhabdus sp. PB62.4, changing incorporation of phenylalanine to leucine alone. Elsewhere, a fifth A-domain fragment was inserted into the middle of a third A-domain and shifted production toward a different major peptide. A five-modular enzyme produced rare cyclic tetrapeptides despite retaining five modules. Disabling its terminal T and thioesterase domains did not eliminate the product, while replacing the recombined C4 domain with an LCL-C2 domain changed the output to cyclic pentapeptides. These results showed that the recombined C4 domain enabled early release and cyclization of tetrapeptides.

Across the family, the researchers predicted at least 12 new main products and detected 49 compounds, including side products, during the experimental analyses. They identified ten unique GxpS variants with product spectra expected to differ from the canonical ortholog. Isotopic feeding and peptide co-injection supported the proposed structures and substrate incorporation. They also found that every tested phenylalanine-incorporating synthetase could incorporate p-aminophenylalanine or its methylated derivative; in almost all feeding tests, the resulting peptide was dominant under the conditions used. For Xenorhabdus species lacking the operon, the phenylalanine-containing GameXPeptide was the main product. The scale of the recovered family diversity therefore extended beyond the canonical peptide, even though the study examined synthetases from only two bacterial genera.

Ancestral sequence reconstruction offered another view of the diversification. The researchers reconstructed the last common ancestor of the analyzed GxpS family, assembled its domains into a complete synthetase and expressed it in Escherichia coli. The inferred ancestor produced a cyclic pentapeptide with a structure similar to that of the canonical modern peptide. A deliberately less probable reconstruction, differing at 244 amino-acid positions, produced the same major product. Mapping peptide structures onto the TTE-domain tree revealed extensive convergence: phenylalanine at the third position was independently replaced by leucine seven times, producing four synthetases with the same major peptide, while 12 independent recombinations introduced valine-specific domains and yielded five origins of another pentapeptide. Tetrapeptide production arose through several routes, including two independently evolved four-modular synthetases and five-modular enzymes that acquired an early-cyclization domain.

Laboratory feeding tests then showed that the structural differences among GXPs translated into different insecticidal activities. The researchers tested eight peptides against larvae of Spodoptera exigua, Plutella xylostella, Tenebrio molitor and Frankliniella occidentalis. Each treatment contained 10 larvae, each was conducted with three biological replicates, and mortality was recorded three days later. Comparison of five models by the Akaike information criterion favored the most complex model, which assigned a distinct potency to each peptide–insect combination. Peptide 36 was the most active against S. exigua but among the least effective against P. xylostella, whereas peptide 4 was most effective against P. xylostella. The experiment nevertheless covered only four insect species, and the authors cautioned against assuming that any one GXP exclusively targets a particular insect in nature.

Recombination statistics supplied evidence that selection affected some events. The researchers distinguished homotypic exchanges, which involved domains with the same substrate specificity and did not change the peptide’s composition, from heterotypic exchanges involving different specificities. Homotypic events served as a proxy for the neutral process. Heterotypic recombinations affected the specificity pocket of the A-domain significantly more often than expected from that proxy; Fisher’s exact tests gave P less than 0.05 in the evaluated tables. This pattern was consistent with rare, specificity-changing events being preferentially retained by selection. It was not a direct measurement of the fitness of ancient peptides, however, and the authors also described a possible non-selective explanation. The 27 synonymous recombination events identified, compared with 26 non-synonymous events, suggested that drift also played a substantial role in GxpS diversification. Likewise, the presence of several peptides with similar potency against each tested insect left room for chemically distinct variants to persist without a large measured activity difference.

A third influence emerged from donor bias. The actual historical A-domain donors were generally more similar at the nucleotide level to their recipients than alternative donors with the same specificity located elsewhere in the same genome. Candidate donors from other non-ribosomal peptide synthetase genes were significantly less similar, with P less than 10⁻⁴ in the relevant paired comparisons. Most A-domain donors also came from within GxpS itself. This pattern suggests a self-reinforcing restriction: recombination among similar domains within the gene would homogenize portions of the sequence and make later exchanges among those internal domains still more likely, while less-similar external domains would remain less accessible. The observed convergence may therefore reflect both the benefits of a particular peptide and the limited set of substitutions that biased recombination made readily available.

Together, the results portray GxpS evolution as the outcome of interacting forces rather than a purely adaptive or purely random process. Selection was consistent with preferential retention of some specificity-changing recombinations, drift accompanied many neutral and nearly neutral structural changes, and recombination bias restricted which amino-acid specificities could be accessed easily. The study also has clear limits. It examined one synthetase family, tested only eight peptides and four insect species, and used homotypic recombination as an approximation to neutrality. Synonymous events were harder to verify experimentally, while some could still have affected peptide production or enzyme folding. The authors further noted that every recombination overwrites part of the sequence record, meaning that some ancient history may be unrecoverable. Whether the same combination of forces shapes other non-ribosomal peptide families remains an open question that will require analyses of other synthetase systems.

Subject of Research: Evolutionary drivers of non-ribosomal peptide diversification in entomopathogenic bacteria

Article Title: Drift, selection and recombination bias drive the evolution of a non-ribosomal peptide

Article References: Kaiser, S., Shahmohammadi, N., Goel, T. S. M., Buchanan, C. J., Garg, S. G., Vidaurre-Barahona, D., Qian, Z., Lefèvre, C., Schindler, D., Kim, Y., Bode, H. B., & Hochberg, G. K. A. (2026). Drift, selection and recombination bias drive the evolution of a non-ribosomal peptide. Nature Ecology & Evolution. https://doi.org/10.1038/s41559-026-03214-y

Image Credits: AI Generated

DOI: 10.1038/s41559-026-03214-y

Keywords: non-ribosomal peptides, GameXPeptides, GxpS, bacterial evolution, genetic recombination, natural selection, genetic drift, convergent evolution, entomopathogenic bacteria, insecticidal peptides, Drift, selection

Researchers phenotyped the maize panel for 34 agronomic traits and used artificial neural network analysis to identify the traits regarded as the most influential contributors to biological yield. Eight traits were selected under optimal phosphorus conditions and five under phosphorus-deficient conditions. The panel was then genotyped with high-density SNP markers, and genomic estimated breeding values were predicted using seven models: Bayes A, Bayes B, Bayes C, Bayesian Lasso, Bayesian Ridge Regression, Genomic Best Linear Unbiased Prediction and Ridge Regression Best Linear Unbiased Prediction.

Phenotypic–genotypic correlations ranged from 0.85 to 0.96 under optimal phosphorus conditions and from 0.61 to 0.95 under phosphorus deficiency. Heritability estimates for the assessed traits were described as moderate to high, although individual estimates were not given in the abstract. Principal component analysis accounted for more than 99% of the total variance, which the authors interpreted as indicating substantial genetic diversity and strong model convergence. Cluster analysis classified the 93 genotypes into three groups. Four genotypes—Ma010, Ma001, Ma011 and Ma013—consistently exhibited favorable genomic estimated breeding values under both phosphorus regimes.

The available evidence supports the relative performance of the evaluated prediction models and identifies promising genotypes within this panel, but it does not demonstrate realized gains in biological yield. The abstract does not identify the trial location, growing conditions, sample sizes for individual traits or performance of each model for each trait. It also does not report an independent validation population, so the stability of the model comparisons and genotype rankings across additional sites, seasons or environments remains unestablished. The authors describe genomic selection combined with statistical and machine-learning analysis as a potential framework for developing maize cultivars adapted to phosphorus-deficient environments; the reported study itself assessed prediction performance rather than the success of a deployed breeding program.

Subject of Research: Genomic prediction of maize traits associated with biological yield under optimal and deficient phosphorus supply

Article Title: Genomic selection for key traits affecting biological yield of maize (Zea mays L.) under optimal phosphorus and phosphorus deficient conditions

Article References: Farhadi, F., Karimzadeh, G., Darvishzadeh, R., Alipour, H., & Razmjoo, M. (2026). Genomic selection for key traits affecting biological yield of maize (Zea mays L.) under optimal phosphorus and phosphorus deficient conditions. BMC Plant Biology. https://doi.org/10.1186/s12870-026-10005-x

Image Credits: AI Generated

DOI: 10.1186/s12870-026-10005-x

Keywords: maize, genomic selection, phosphorus deficiency, genomic prediction, Bayesian models, heritability, biological yield, high-density SNP markers, maize genotypes, Genomic, selection, traits

Bt maize is a genetically modified crop that produces insecticidal proteins derived from the bacterium Bacillus thuringiensis . The proteins act against susceptible herbivorous pests, including lepidopteran insects. Wind-dispersed maize pollen can settle on larval host plants, where caterpillars may ingest Bt proteins and experience additional mortality or sublethal adverse effects.

European Union rules require environmental risk assessment of genetically modified organisms before market placement, and controlled laboratory tests form a central part of assessing potential hazards to non-target organisms. Because thousands of European butterfly and moth species exist, however, it is not feasible to test every species that could be exposed. The researchers began with the LEPIDAT database of the German Federal Agency for Nature Conservation, which contained 8,670 entries for lepidopteran taxonomic units, including genera, species and subspecies.

A preliminary screen retained 1,478 entries for which information was available on altitude, habitat, larval feeding period and feeding type. Most “Micro-Lepidoptera” were excluded before the formal selection because their biological information was insufficient, although the diamondback moth, Plutella xylostella , was retained as an exception. The researchers regarded it as an abundant farmland moth, sensitive to Bt, and easy to breed in the laboratory.

The first formal step excluded 350 listed species in the family Noctuidae, some of which are now classified in Nolidae, because members of this group had been reported to be comparatively insensitive to Bt toxins. The next steps removed species that did not occur in Germany and those unlikely to be exposed to Bt maize pollen. The latter exclusions covered species restricted to elevations above 700 metres, where maize is not cultivated in Germany; species whose larval periods did not overlap with the approximate German maize pollen-shedding period from late June to late August; and species whose larvae feed within host plants or below ground on roots, where deposited pollen would not normally reach them.

These filters reduced the candidate pool to 652 potentially exposed species. The remaining species were classified into broad habitat types: farmland, dryland, wetland, woodland and settlements. A species could qualify for more than one habitat category.

The researchers selected candidates found concurrently in at least four habitat types, including farmland, to favor widespread species capable of representing a range of landscapes surrounding maize cultivation. This stage retained 54 species. Laboratory practicality then provided the next filter.

Candidate species needed evidence that females could lay eggs under laboratory conditions and that their larvae could be reared indoors on host-plant leaves or an artificial diet. Twenty-two species without such evidence were removed, leaving 32 species considered suitable for laboratory experiments. Conservation status and expert judgment shaped the final choice.

Protected species were ranked highly because risk assessment must consider potential harm to populations that are conservation targets under European legislation. The researchers then assessed taxonomic variety, the geographic and ecological range represented by the candidates, and differences in larval body size. These criteria produced a tentative panel of 15 focal species from four lepidopteran superfamilies.

Breeding feasibility was supported by a separate literature review covering nine superfamilies and 30 families of European Lepidoptera. The search identified 548 publications, including handbooks, reviews, traditional breeding guides, modern rearing studies and artificial-diet protocols. Some records dated from more than a century ago, while others described controlled methods and detailed environmental requirements.

Although the review found evidence for laboratory culture across a broad taxonomic range, it also showed that an individual species may require time and adaptation before a reliable method can produce enough healthy larvae for testing. Natural host plants most closely reproduce the circumstances in which caterpillars would encounter deposited maize pollen, and the source study notes that they are the recommended approach for Bt maize effect tests. Artificial diets can nevertheless aid standardization, reproducibility or rapid screening.

Common formulations contain flour from legumes or cereals, protein, yeast, cholesterol, vitamins, minerals and agar as a binder; some also include dried plant material or sucrose to stimulate feeding. Such diets are not universally accepted, especially by newly hatched larvae, and larvae that have already fed on natural food may not transfer successfully. Artificial food must be replaced before it dries or develops mould, and the cited guidance advises against allowing larvae to pupate within it.

Laboratory culture also requires control of several biological and physical factors, including host-plant identity and availability, temperature, humidity, photoperiod, diapause, mating conditions, oviposition substrates and population density. Regular replacement of food, adequate ventilation and routine disinfection of containers and handling equipment help limit condensation, mould and disease. Colonies maintained over several generations may require measures to avoid inbreeding and, in some cases, supplementation with field-collected individuals.

Eggs or pupae may be disinfected before mass rearing, while field-collected food plants can introduce pathogens or parasitoid eggs. These practical requirements are why the researchers treated successful breeding as a central selection criterion rather than a detail to be addressed only after choosing test organisms. The procedure is presented as a selection protocol, not as evidence that particular butterflies or moths are hazardous or protected from Bt exposure.

Susceptibility to Bt maize pollen remains poorly known for most non-target lepidopterans, which is why it was not used as the main screening factor. The final two stages also relied on expert assessment, although that judgment was applied within the documented sequence. The German list cannot be transferred unchanged to other regions: researchers would need to adapt altitude limits, pollen-shedding periods, habitat information and the species present in the relevant receiving environment.

Some excluded species may also deserve testing because information was missing rather than because they were unsuitable. Even with those limitations, the framework offers a systematic way to turn a large regional fauna into a diverse, ecologically informed and practically testable set of candidate non-target insects.

Subject of Research: Selection of non-target butterflies and moths for laboratory assessment of Bt maize pollen effects.

Article Title: Selection of non-target Lepidoptera species to test Bt maize effects in the laboratory: which species and how to breed them?

Article References: Lang, A., Dolek, M., Lee, M. S., Freese-Hager, A., & Otto, M. (2020). Selection of non-target Lepidoptera species to test Bt maize effects in the laboratory: which species and how to breed them?. BioRisk, 15, 45-65. https://doi.org/10.3897/biorisk.15.59823

Image Credits: AI Generated

DOI: 10.3897/biorisk.15.59823

Keywords: Bt maize, non-target Lepidoptera, environmental risk assessment, laboratory breeding, maize pollen, butterflies, moths, species selection, insect conservation, artificial diets, Selection, non-target

The study was designed as a retrospective cohort analysis drawing on two of the most widely used critical care data repositories in the world. The MIMIC-IV database, which contains fully de-identified records of patients treated at Beth Israel Deaconess Medical Center, supplied the training and internal validation data, while the eICU Collaborative Research Database, a multicenter collection of ICU stays across the United States, served as an independent external validation cohort. The primary outcome was unambiguous: death from any cause within 28 days of ICU admission. Because the underlying data were de-identified, the Institutional Review Board of Beth Israel Deaconess Medical Center and MIT waived the need for patient informed consent, and all analytical procedures followed the MIMIC-IV Data Use Agreement. This dual-database architecture is a deliberate methodological choice, forcing any candidate model to prove itself on patient populations it has never seen before rather than merely fitting the data it was trained on.

Before a single algorithm was trained, the team confronted the messy realities of clinical data. Variables that exhibit multicollinearity, meaning they carry overlapping statistical information, can distort model coefficients and inflate false confidence in individual predictors. The researchers eliminated such redundancy using Generalized Variance Inflation Factors, a technique that quantifies how much the variance of an estimated regression coefficient is amplified by correlation among predictors. To select the most robust prognostic features without leaking information from the validation sets into the training process, the team deployed a nested ten-fold cross-validation framework that jointly combined three feature selection algorithms: LASSO, a regression method that shrinks irrelevant coefficients to zero; Random Forest, an ensemble of decision trees whose variable importance rankings highlight genuinely informative features; and a Genetic Algorithm, an evolutionary optimization procedure that searches for the feature subset that maximizes predictive performance. Only features that survived this triple interrogation were carried forward.

The result of that joint selection was a compact panel of 14 core prognostic features. Among them were the Simplified Acute Physiology Score II, a well-established severity-of-illness index computed from physiological measurements; Red Cell Distribution Width, a routine blood test marker that reflects variability in red blood cell size and has emerged as a surprisingly powerful predictor of critical illness outcomes; the Charlson Comorbidity Index, which sums the mortality burden of a patient’s chronic conditions; and serum lactate, a biochemical signature of tissue hypoperfusion and shock. The convergence of three fundamentally different selection algorithms on the same feature panel is itself a form of validation, suggesting these variables capture real biological and clinical signal rather than statistical noise particular to one modeling approach.

With the feature panel fixed, the researchers systematically evaluated 20 different machine learning algorithms under a stratified five-times repeated ten-fold cross-validation framework, reinforced by 1000-iteration Bootstrap resampling to quantify uncertainty in every performance metric. Critically, they preserved the natural, real-world class distribution of the data rather than artificially balancing the mortality and survival classes, a decision that keeps the model honest about the base rates it will encounter in practice. Among the 20 contenders, LightGBM, a gradient boosting framework that builds decision trees sequentially and uses histogram-based splitting for remarkable computational efficiency, demonstrated the most outstanding overall performance. In internal testing, the model achieved an Area Under the Receiver Operating Characteristic Curve of 0.8253, with a 95% confidence interval of 0.7824 to 0.8648, alongside an accuracy of 0.8434, a specificity of 0.9745, and a Brier Score of 0.1106, indicating strong probability calibration. Decision Curve Analysis further showed significant net clinical benefit, meaning the model would improve decisions across a clinically meaningful range of risk thresholds.

But the story took a sobering turn when the model faced the independent external validation set from the eICU database. That cohort was characterized by a small sample size and extreme class imbalance, with far fewer deaths than survivals, and under those conditions the model’s predictive efficacy was significantly attenuated. The AUC dropped to 0.615, with a 95% confidence interval of 0.413 to 0.817, revealing notable negative bias and a tendency toward risk overestimation. This is not a footnote; it is the central methodological lesson of the study. A model can look brilliant on the database it was trained on and still stumble when the patient population, documentation practices, and case mix shift. The authors are candid about this limitation, noting that the small sample size of the external cohort posed constraints on assessing the model’s generalization, and their transparency is a refreshing contrast to the overclaiming that sometimes accompanies machine learning papers in medicine.

What elevates the work beyond raw prediction is its commitment to interpretability. Black-box models have long been a barrier to clinical adoption, and the researchers addressed this by dissecting the optimal model with two complementary frameworks: SHAP, or SHapley Additive exPlanations, which borrows concepts from cooperative game theory to assign each feature a fair contribution to every individual prediction, and LIME, or Local Interpretable Model-agnostic Explanations, which fits simple local surrogate models around individual predictions to reveal which factors pushed a specific patient’s risk up or down. Together, SHAP and LIME provide visual interpretations of the prognostic features, allowing a clinician to see not just that a patient is high risk but why, whether it is a rising lactate, a widened RDW, or a heavy comorbidity burden driving the estimate.

The team also translated the model into practice by deploying it as a Streamlit-based interactive web application, freely accessible online, that computes an individualized 28-day mortality risk estimate from a patient’s admission features. The tool is designed for early risk stratification at the bedside, enabling frontline physicians to identify patients who warrant escalated monitoring, earlier palliative care conversations, or more aggressive interventions, and to ground those decisions in robust, evidence-based support rather than intuition alone. The web deployment also serves as a kind of public stress test, inviting scrutiny of how the model behaves on inputs drawn from centers beyond its training data.

For the field of predictive medicine, the study delivers a dual message. On one hand, it demonstrates that a carefully engineered pipeline, combining rigorous collinearity screening, nested multi-algorithm feature selection, and honest evaluation under natural class distributions, can produce a genuinely useful prognostic instrument for one of the most vulnerable patient populations in critical care. On the other hand, it shows that external validation remains the gatekeeper of clinical credibility, and that a model’s performance can attenuate dramatically when cohort characteristics shift. The 14-feature panel, anchored by severity scores, hematologic markers, and metabolic signals, now stands as a candidate reference standard for 28-day mortality prediction in critically ill colorectal cancer patients, pending larger external cohorts. As machine learning continues to seep into the ICU, this study offers both a blueprint for doing it right and a warning about the gap between internal brilliance and real-world reliability.

Subject of Research: Machine learning prognosis in critical care oncology

Article Title: Multi-algorithm joint feature selection for 28-day prognosis in critically ill patients with colorectal cancer: lightGBM-based model development and multicenter validation

Article References: Wang, W., Li, M., & Shen, T. (2026). Multi-algorithm joint feature selection for 28-day prognosis in critically ill patients with colorectal cancer: lightGBM-based model development and multicenter validation. BMC Cancer. https://doi.org/10.1186/s12885-026-16947-7

Image Credits: AI Generated

DOI: 10.1186/s12885-026-16947-7

Keywords: machine learning, colorectal cancer, ICU, LightGBM, prognosis, predictive medicine, Multi-algorithm, joint, feature, selection, critically, patients

The research, led by Martina Mercurio and Concetta Rispoli of the University of Naples Federico II together with colleagues from several Italian institutions, took a fresh look at the pozzolana deposits of the Campi Flegrei, the volcanic district west of Naples. The team collected samples from historical quarries that the Romans themselves are known to have exploited, extending from Naples to Bacoli. One sample, labeled SAV, came from an ancient tuff quarry in Marano di Napoli and represents pozzolana in the strict sense, the unconsolidated facies of the Neapolitan Yellow Tuff. Two others, labeled CL and LP, were collected near Baia, close to the Pontile Lubrano, an area where Imperial-age ceramics and artifacts testify to intense Roman maritime activity. Crucially, these two deposits belong to eruptions that occurred after the Neapolitan Yellow Tuff event, within the caldera’s later Epoch II activity.

The geological setting explains why this region is so rich in reactive ash. The Campi Flegrei caldera took shape through two cataclysmic collapses: the Campanian Ignimbrite eruption about 40,000 years ago, which ejected between roughly 60 and 300 cubic kilometers of phonolitic-trachytic material and ranks as the largest eruption in the Mediterranean over the past 200,000 years, and the Neapolitan Yellow Tuff eruption about 15,000 years ago, which produced around 50 cubic kilometers. After that second caldera-forming event, more than 70 eruptions occurred within the caldera, grouped into three epochs, the most recent ending with the Monte Nuovo eruption in AD 1538. These later eruptions, mostly explosive events of intermediate to Plinian scale, scattered cones and ash deposits across the entire territory. The magmas are potassic-alkaline in character, dominated by trachytic and trachy-phonolitic compositions, with clinopyroxene, plagioclase, sanidine, biotite and magnetite as the principal mineral phases.

To determine whether these post-Neapolitan Yellow Tuff deposits could genuinely serve as hydraulic binders, the researchers deployed a battery of modern analytical techniques. Polarized optical microscopy revealed the texture of pumice fragments and their crystal cargo; X-ray powder diffraction identified the mineral assemblage, including the zeolites phillipsite and analcime that form when hot reactive glass interacts with hydromagmatic fluids; field-emission scanning electron microscopy coupled with energy-dispersive spectroscopy delivered precise chemical analyses of individual volcanic glass fragments; and X-ray fluorescence characterized the whole-rock composition. Laser particle-size analysis measured the grain-size distribution of the fine fractions, the portion Roman craftsmen would have crushed or selected deliberately, since finer particles expose more reactive surface area. The samples proved remarkably similar in texture and mineralogy, all containing fresh volcanic glass, pumice fragments with elongated vesicles, and the zeolitic association typical of Phlegraean pyroclastic products.

The chemistry told the most compelling story. On the standard Total Alkali versus Silica classification diagram, the whole-rock compositions of CL and LP plot as trachyte while SAV falls in the latite field. But the decisive evidence came from individual point analyses of the glass component itself, which bypasses the diluting influence of crystals, lithic fragments and alteration products. Glass from CL and LP overlaps the compositional fields of the Baia-Fondi di Baia eruption, a small multievent eruption that occurred 9,500 to 9,600 years ago on the western side of Pozzuoli, while SAV glass matches the Neapolitan Yellow Tuff field. The potassium-to-sodium ratio proved a particularly sharp discriminator: CL and LP glasses show ratios of two or higher, excluding the low-ratio Averno eruption, whereas SAV aligns with the Neapolitan Yellow Tuff. These geochemical attributions agree with the official geological cartography of the sampling sites.

The team then confronted their quarry samples with published glass analyses of pozzolanic aggregates extracted from actual Roman mortars across the Mediterranean. The dataset spanned local Campanian monuments, including the Baths of Baia, the Flavian Amphitheater of Pozzuoli and the vast Piscina Mirabilis cistern, as well as distal sites: Roman structures in the Venice Lagoon, the Roman theater of Aquileia in northeastern Italy and the ancient city of Hippo Regius in Algeria. The glass compositions from these mortars fall within the compositional range of the Neapolitan Yellow Tuff and, tellingly, match the variability of Baia-Fondi di Baia products as well. Some anomalies in calcium oxide and sodium oxide values for the Venice and Aquileia samples reflect severe post-depositional alteration in underwater environments rather than original compositions. The pattern suggests that Roman builders did not confine themselves to a single eruption but exploited multiple Phlegraean deposits as pulvis puteolana.

A machine-learning tool added statistical rigor to the visual comparisons. Discriminant analysis, a supervised multivariate method suited to datasets with overlapping chemical features, assigned the quarry samples and mortar aggregates to specific eruptions with high probabilities. CL samples matched the Baia-Fondi di Baia eruption with probabilities from about 93 to 100 percent, LP from nearly 100 percent, and SAV, despite its geological mapping to the Neapolitan Yellow Tuff, was classified as Baia-Fondi di Baia with 95.6 to 100 percent probability, reflecting the genuine chemical overlap shared by several Phlegraean products. The literature mortar data showed primary affinity with Baia-Fondi di Baia followed by the Neapolitan Yellow Tuff, with a minor subset pointing to Monte Spina and Averno. The authors note that these probabilistic attributions rest on major-element chemistry alone and that trace-element data could sharpen future provenance assignments.

The technological test came from the Fratini test, a European standard method developed more than fifty years ago that measures how much calcium hydroxide a blended cement can consume over fifteen days of curing at 40 degrees Celsius. All three samples plotted below the calcium hydroxide solubility curve, proving genuine pozzolanic activity. The SAV sample showed the highest calcium oxide reduction at 38.6 percent, followed by CL at 30.3 percent and LP at 29.2 percent. The authors are careful to frame these results as complementary evidence: the Fratini test was designed for modern cementitious systems, and the hypothesis that Romans deliberately selected specific Phlegraean materials rests mainly on the convergence of historical quarry locations, petrographic features, reactive glass content and the geochemical match between deposits and mortar aggregates.

Why did these particular ashes work so well? The answer lies in the polymerization of the volcanic glass network. Glasses with moderate silica content, characteristic of part of the Neapolitan Yellow Tuff and Baia-Fondi di Baia products, are less polymerized and therefore dissolve more readily, accelerating the pozzolanic reaction in which silica and alumina react with calcium hydroxide at room temperature to form calcium-aluminum-silicate-hydrate gel. This C-A-S-H gel is what gives Roman hydraulic mortars their ability to harden underwater and their exceptional mechanical strength. Silica-rich glasses, by contrast, are less reactive. The shared compositional variability of the two eruptions in terms of silica content may thus be the key factor behind high-performance hydraulic mortars, and it may explain why the Romans favored these deposits without ever knowing the chemistry behind them.

The study establishes a first reference database of glass analyses for Phlegraean pozzolana, offering archaeologists a powerful tool for tracing the provenance of mortar aggregates in monuments across the former Roman Empire. It also reframes the scale of Roman resource management: the expansion of Puteoli’s harbor made the ash a strategic commodity, exported along grain-ship ballast routes to Ostia, Alexandria, Caesarea Maritima and beyond, and even carried to Constantinople, as the fifth-century historian Sidonius Apollinaris documented in his Panegyricus. Roman craftsmen, the authors conclude, possessed an advanced empirical knowledge of the raw materials of their territory, recognizing the value of many pyroclastic deposits rather than a single one. That legacy, encoded in mortars that have survived two millennia of seawater, may yet inform modern sustainable cement formulations seeking to rediscover what the Romans knew by practice alone.

Subject of Research: Provenance and selection of Roman hydraulic pozzolana from the Phlegraean Fields

Article Title: On the tracks of Pulvis Puteolana: provenance and selection of Roman hydraulic materials

Article References: Mercurio, M., Montesano, G., Cappelletti, P., Petrosino, P., Graziano, S. F., Campanile, A., Liguori, B., & Rispoli, C. (2026). On the tracks of Pulvis Puteolana: provenance and selection of Roman hydraulic materials. Archaeological and Anthropological Sciences, 18(11), Article 217. https://doi.org/10.1007/s12520-026-02582-7

Image Credits: AI Generated

DOI: 10.1007/s12520-026-02582-7

Keywords: tracks, Pulvis, Puteolana, provenance, selection, Roman, hydraulic, materials, scientific research

The research, led by Miriam Aledda and Achim Kohler at the Norwegian University of Life Sciences in Ås, together with collaborators at the University of Rome La Sapienza, the Norwegian food research institute Nofima, and the University of L’Aquila, addresses a challenge that has grown more pressing as laser technology has advanced. Quantum cascade lasers and interband cascade lasers have emerged over the past decade as key light sources for mid-infrared sensing, offering high intensity, excellent signal-to-noise ratios, and exceptional spectral resolution. Unlike traditional broadband Fourier-transform infrared spectrometers, which collect data across the entire mid-infrared range from 4000 to 400 wavenumbers, these lasers can be engineered to emit precisely in narrow spectral windows. That precision is their greatest strength and their greatest design constraint.

The team focused on a particularly important spectral region: 1700 to 1600 wavenumbers, which corresponds to the amide I band. This band, arising from the stretching vibration of carbon-oxygen double bonds in the peptide backbone of proteins, is exquisitely sensitive to protein secondary structure and conformational changes. It is also the region where a tunable interband cascade laser developed by the German company nanoplus currently operates, capable in principle of emitting at more than 1000 discrete frequencies through temperature and current modulation. The researchers wanted to know: if you can only access a limited number of wavelengths, which ones should you choose?

To answer this, the team assembled three datasets of protein hydrolysates derived from salmon and poultry by-products, materials obtained through routine food industrial processes. Protein hydrolysates are produced by enzymatically breaking down protein-rich raw materials into smaller peptides and free amino acids, and their average molecular weight is a crucial quality parameter that determines functional properties, bioavailability, and sensory characteristics. The datasets comprised 95 salmon hydrolysate samples, 133 poultry samples measured in five technical replicates, and 286 poultry samples measured in three technical replicates, with average molecular weights ranging from roughly 1885 to 7111 grams per mole as determined by size-exclusion chromatography. The researchers then applied a battery of variable selection methods to identify which wavelengths in the mid-infrared spectrum carry the most predictive information for estimating average molecular weight.

The methods tested span several families of machine learning approaches. Sparse Partial Least Squares, an embedded technique that performs variable selection and dimensionality reduction simultaneously, was applied with varying degrees of sparsity tailored to different spectral ranges. The Successive Projection Algorithm, a forward selection method designed to minimise collinearity among selected variables, was also included. The Covariance Selection family, comprising the original CovSel, its group-based extension g-CovSel, and a newly implemented window-based variant called w-CovSel, formed another major component of the comparison. Principal Variable Selection and its supervised extension Principal Variable Regression rounded out the set. Each method operates on different principles, from maximising covariance with the response variable to selecting groups or narrow windows of consecutive variables, and each carries different implications for laser hardware design.

The results revealed a striking hierarchy of predictive power across spectral regions. When the researchers benchmarked performance using broadband partial least squares regression on the full mid-infrared range, they obtained coefficients of determination of 0.78 for the fish dataset, 0.85 for the first poultry dataset, and 0.84 for the second. Narrowing to the fingerprint region from 1800 to 800 wavenumbers improved these values for two of the three datasets, and restricting further to the amide I range from 1700 to 1600 wavenumbers maintained high performance, with the second poultry dataset actually improving to 0.90. Remarkably, even randomly selecting ten wavelengths within the amide I range produced coefficients of determination of 0.79, 0.75, and 0.84 across the three datasets, approaching the performance of broadband models. This finding underscores how information-dense the amide I region is for protein characterisation.

Perhaps the most surprising result concerned the set of biochemically motivated features drawn from the literature. The researchers identified nine wavenumbers that prior studies had flagged as particularly relevant for protein and peptide characterisation, including signals at 1740, 1645, 1630, 1548, 1515, 1452, 1400, 1081, and 1045 wavenumbers. When these were used alone for regression, performance dropped considerably, with coefficients of determination of only 0.58, 0.74, and 0.62 across the three datasets. This suggests that variable selection based purely on literature-derived biochemical knowledge is not the best approach for laser design, especially when the number of accessible wavelengths is limited. The data-driven methods, by contrast, consistently identified wavelengths that outperformed the literature-based selection.

Among the data-driven approaches, the group-based and window-based Covariance Selection methods proved particularly suited to laser design constraints. The g-CovSel method, which selects groups of variables rather than individual wavelengths, achieved the best overall performance on the fish dataset with a coefficient of determination of 0.87 when all five selected groups were combined. The newly proposed w-CovSel method, designed specifically to select narrow windows of three consecutive variables, mimics the fine tuning capability of interband cascade lasers, which can achieve sub-wavenumber precision around discrete emission points. When the researchers applied w-CovSel within the amide I range, selecting up to five narrow intervals, prediction ability was very good, consistent with previous findings that three fine-tuned single emission points already yield strong results. The team also found that the first interval selected by w-CovSel within the amide I range for the second poultry dataset alone achieved a coefficient of determination of 0.84, demonstrating that even a single narrow tuning window can carry substantial predictive information.

The implications extend beyond protein hydrolysates. The researchers note that the same variable selection strategies could inform the design of other advanced optical systems, including hyperspectral imaging platforms, Raman spectroscopy setups, and atomic force microscopy-based spectroscopic tools, wherever spectral sparsity, tuning precision, and data interpretability are critical. As tunable laser technology continues to mature, the ability to systematically identify which wavelengths matter most for a given analytical task will be essential for building compact, high-performance instruments that bypass the optical complexity of traditional broadband spectrometers. The study demonstrates that machine learning is not merely a tool for analysing spectroscopic data after the fact, but a genuine design partner for the engineers building the next generation of mid-infrared laser instrumentation.

Subject of Research: Wavelength selection for laser design in mid-infrared spectroscopy

Article Title: Wavelength selection for laser design in mid-infrared spectroscopy

Article References: Aledda, M., Marini, F., Kafle, B., Erdem, M. C., Karki, P., Liland, K. H., Zimmermann, B., Afseth, N. K., Biancolillo, A., Tafintseva, V., Tøndel, K., Shapaval, V., & Kohler, A. (2026). Wavelength selection for laser design in mid-infrared spectroscopy. Frontiers in Photonics, 7, Article 1696425. https://doi.org/10.3389/fphot.2026.1696425

Image Credits: AI Generated

DOI: 10.3389/fphot.2026.1696425

Keywords: Wavelength, selection, laser, design, mid-infrared, spectroscopy, scientific research

The research team trained Decision Tree, Random Forest, and XGBoost models on two years of historical one-minute data spanning January 2022 to December 2023. Rather than estimating exact prices, the models were reframed as classification problems, predicting whether future price movements would exceed the asset’s transaction cost within forecasting horizons of 1, 3, 5, 10, 15, and 30 minutes. This approach provides a probabilistic estimate of trend direction rather than a point forecast, which has shown superior results in previous literature. The predictive variables were intentionally restricted to technical indicators derived from high-frequency price and volume data, including moving averages, VWAP, MACD, the Stochastic Oscillator, Bollinger Bands, and the Relative Strength Index, deliberately excluding fundamental variables such as earnings reports or macroeconomic indicators that update at substantially lower frequencies.

One of the most striking findings concerns the forecasting horizon. Across the majority of the 40 analyzed assets, the 30-minute horizon emerged as the best-performing option within the evaluated candidate set. Among the 24 stocks, 22 selected the 30-minute horizon, while all 10 foreign exchange pairs favored this configuration, and 4 of the 6 cryptocurrencies also preferred it. The researchers caution that because 30 minutes represents the upper boundary of the candidate set, this result should not be interpreted as establishing a globally optimal forecasting horizon. Rather, it indicates a tendency toward improved performance at longer horizons, suggesting that immediate price fluctuations at the one-minute scale contain substantial short-term noise that longer intervals can filter out.

The methodology employed a walk-forward rolling-window validation framework with Bayesian hyperparameter optimization, ensuring that all reported test predictions were strictly out of sample. Class imbalance, a common challenge in intraday data where small price movements are more frequent than extreme ones, was addressed through class weighting optimized jointly with the remaining hyperparameters. Notably, rather than optimizing standard accuracy metrics, the researchers conducted Bayesian optimization using the F-beta Score with beta equal to 0.10, a metric that prioritizes precision over recall. This design choice reflects the economic reality that the cost of false-positive trading signals is high, where a small number of adverse trades can offset a larger number of profitable ones.

Regarding model selection, XGBoost was the most frequently selected learning algorithm, chosen for 18 of the 40 analyzed assets, followed by Decision Tree for 13 assets and Random Forest for 9. However, the preferred model varied across asset classes. Random Forest was the most frequently selected model for cryptocurrencies, accounting for 3 of the 6 assets, while XGBoost dominated among foreign exchange pairs. This behavior suggests that no single learning algorithm is universally preferable for all financial instruments, reinforcing the importance of evaluating multiple model families when developing intraday forecasting strategies.

The predictive results revealed substantial heterogeneity across asset classes. Stocks exhibited the strongest overall predictive performance, with an average F-beta Score of 43.88 percent, concentrated primarily in the Information Technology sector. Cryptocurrencies achieved competitive predictive performance, particularly among major cryptocurrencies, with Bitcoin and Ethereum exhibiting precision values exceeding 50 percent. Foreign exchange pairs remained more difficult to predict, with exotic currency pairs showing markedly weaker results than major pairs. Yet the researchers emphasize that superior predictive performance did not systematically translate into economic profitability.

After incorporating realistic transaction costs, the economic picture became considerably more restrictive. Only 5 of the 40 analyzed assets generated a positive average return per trade, while 6 assets achieved positive cumulative profits over the two-year evaluation period. Among stocks, only Tesla, Apple, and Microsoft generated both positive average returns and positive cumulative profits, with Tesla exhibiting the strongest risk-adjusted performance as measured by the Sharpe and Calmar ratios. None of the analyzed currency pairs generated positive economic returns after transaction costs, providing a clear illustration of the distinction between predictive and economic performance. Cryptocurrencies exhibited the strongest economic results, with Bitcoin, Ethereum, and BNB producing the most favorable outcomes.

The researchers also introduced the Maximum Tolerable Slippage, a novel metric quantifying the maximum average adverse price movement per execution that each strategy can tolerate while preserving positive net profitability. Tesla exhibited the highest MTS among equities, indicating greater robustness to execution costs, while Bitcoin showed comparatively high tolerance among cryptocurrencies. Benchmark comparisons against buy-and-hold, naïve persistence-based, and random-direction strategies further demonstrated that the profitable machine learning strategies consistently outperformed random trading, providing evidence that the learned predictive signals contributed meaningfully to the observed economic performance.

The study’s conclusions carry important implications for both academic research and practical trading. The findings suggest that economically viable strategies emerge from the combined selection of an appropriate forecasting horizon, learning algorithm, and prediction target rather than from any individual model alone. The researchers note that predictive accuracy alone is insufficient to assess the practical value of a trading strategy, and that profitability should be evaluated jointly with financial risk, benchmark comparisons, and execution-related considerations. As machine learning continues to attract attention in quantitative finance, this work provides a comprehensive framework for jointly assessing predictive and economic performance under realistic market conditions, reminding us that the market’s efficiency, while perhaps not absolute, remains a formidable barrier to consistent abnormal profits.

Subject of Research: Computational Finance

Article Title: Intraday trading strategy selection under transaction costs with machine learning

Article References: Gaona, A., Rangel-Madariaga, J., & Uribe, L. (2026). Intraday trading strategy selection under transaction costs with machine learning. Frontiers in Applied Mathematics and Statistics, 12, Article 1888767. https://doi.org/10.3389/fams.2026.1888767

Image Credits: AI Generated

DOI: 10.3389/fams.2026.1888767

Keywords: machine learning, intraday trading, transaction costs, XGBoost, Random Forest, cryptocurrency, financial forecasting, Intraday, trading, strategy, selection, under

The problem is deceptively simple. Studies of post-encroachment time typically rely on observed values extracted from video analysis, which limits the sample to vehicles that engaged in at least one interaction with another vehicle or vulnerable road user. Vehicles that pass through the intersection without any observable engagement are excluded from the analysis. Yet these non-engaging vehicles may represent a fundamentally different portion of the driving population—one whose behavior under higher traffic volumes or platooned arrivals could differ markedly from those that did engage. The counterfactual question—what would the post-encroachment time have been for a vehicle that never engaged?—is the critical one, because it would present a full description of the true underlying propensity for dangerous interactions in terms of both duration and frequency.

This is not a new concern in traffic safety research. Prior studies have raised the alarm about selection bias in crash severity modeling, noting that riskier drivers can be over-represented in reported crash data, leading to biased and inconsistent parameter estimates. Similar selectivity problems have been documented during the COVID-19 pandemic, where people who chose to engage in social activities carried an underlying risk proneness that also predisposed them to riskier driving. But the Texas Tech team, led by Karam Daoud, Abdallah Alrawashdeh, and Narayan Venkataraman, is among the first to tackle the issue in the context of post-encroachment time engagements at signalized intersections using a full-information maximum likelihood framework.

The conventional tool for correcting selection bias is the Heckman two-step approach, where the probability of an engagement or crash is factored into a second-stage model. But the Heckman technique relies on the Inverse Mills ratio, which provides meaningful results only in a linear regression setting. Its applicability to nonlinear settings—such as the ordered probit models needed for severity analysis—is questionable. The full-information maximum likelihood approach, by contrast, jointly models the selection and outcome equations, measuring the degree of selection directly through a correlation parameter between the error terms of the two equations. This correlation arises from unobservables shared between the selection process and the outcome process—factors like car-following behavior in downtown environments or pedestrian crossing tendencies that influence both the probability of engagement and its severity.

The computational complexity of the full-information maximum likelihood approach has historically limited its use, especially in larger datasets. The Texas Tech researchers overcame this by carefully scaling variables and re-parameterizing the correlation parameter on the real line, which supported stable and smooth convergence without the need to switch optimization techniques mid-estimation. They transformed the correlation parameter through the inverse hyperbolic tangent function, expanding the parameter space to enable smoother likelihood behavior. The result was a convergent joint model applied to 82,666 observations extracted from the Bellevue intersection, with 30,185 observations yielding measured post-encroachment times and the remaining 52,481 contributing to the selection analysis.

The severity outcome was modeled as an ordered probit with three levels: interaction (post-encroachment time greater than 3 seconds), minor engagement (greater than 2 seconds and less than or equal to 3 seconds), and critical engagement (less than or equal to 2 seconds). These thresholds align closely with the Bellevue Vision Zero conflict-classification framework, which defines post-encroachment time values below 2 seconds as critical and between 2 and 3 seconds as minor. The dataset contained 28,000 observations for interactions, 1,399 for minor engagements, and 786 for critical engagements—a stark imbalance that reflects the reality that most engagements at signalized intersections are low-severity interactions rather than near-crashes.

The joint model converged at a log-likelihood of −54,056.836, compared to −54,061.188 for the ordered probit model without selection. The likelihood-ratio test produced a chi-squared value of 8.704, demonstrating statistically significant rejection of the null hypothesis that there is no correlation due to unobservables. The correlation parameter itself was estimated at −0.174 with a t-statistic of −2.78, confirming that shared unobservables—hypothesized to include driver tendencies unique to the downtown context such as interactions with parked vehicles, non-motorized traffic, gap acceptance behavior, and perceived turning movement safety—exert a real influence on both engagement occurrence and severity.

When the researchers compared coefficients from the joint model to those from the naïve ordered probit without selection, the bias was unmistakable. Standard errors in the outcome equation were on average 14 percent higher under selection. Coefficient bias averaged 5.8 percent, but the largest distortions appeared for the speed ratio variable—measuring the ratio of average to median trajectory speed as a proxy for speed skewness—and the daytime indicator, with changes of 35 percent and 73 percent respectively. The speed ratio effect was over-estimated in magnitude in the presence of selection, while the daytime variable was under-estimated. Signs remained unchanged, but the true direction and magnitude of effects by severity level could only be recovered through a comprehensive analysis of marginal effects.

The marginal effects analysis revealed three types of behavioral insight: direct effects from variables in the severity equation, indirect effects from variables in the selection equation, and full effects combining both. The speed skew measure exerted noteworthy full marginal effects on the probability of the lowest severity level, as did lane width difference and lane count difference between departure and approach legs. The lane count difference variable—the difference between the number of lanes on the departure leg and the approach leg—had the greatest magnitude effect, contributing a −0.449 decrease in the probability of interaction engagement while increasing the probabilities of minor and critical engagements. This suggests that lane balance between approach and departure legs can yield lowered encroachment severity likelihoods, while dedicated turning movement lanes or parking-induced lane reductions may inadvertently escalate engagement severities through merging maneuvers.

The policy implications are concrete. Traffic control measures that mitigate right-skewed speed distributions can yield beneficial effects in terms of lowered post-encroachment times. Narrower approach lane widths to accommodate higher turning movement capacity may support lower encroachment severities. Approach leg medians, sharper turning radii, and departure leg right shoulders all exerted measurable effects on engagement probabilities across the severity spectrum. The researchers recommend extending the 10-second measurement technique to multiple intersections to capture richer behavioral insights across downtown, urban, suburban, and rural contexts, particularly for variables related to traffic operations and control, capacity recapture policies such as elimination of approach leg parking, and lane balance decisions. The study focused on a single intersection, so findings capture within-site variations and selection effects—but the framework itself offers a replicable template for any traffic safety analysis that relies on observed engagements rather than the full population of vehicles passing through a conflict zone.

Subject of Research: Selection bias in post-encroachment time conflict severity modeling using a full information maximum likelihood framework

Article Title: Selection bias in post-encroachment time conflict severity modeling using a full information maximum likelihood framework

Article References: Daoud, K., Alrawashdeh, A., & Venkataraman, N. (2026). Selection bias in post-encroachment time conflict severity modeling using a full information maximum likelihood framework. Frontiers in Future Transportation, 7, Article 1893114. https://doi.org/10.3389/ffutr.2026.1893114

Image Credits: AI Generated

DOI: 10.3389/ffutr.2026.1893114

Keywords: traffic safety, selection bias, post-encroachment time, full-information maximum likelihood, ordered probit, signalized intersections, trajectory analysis, Vision Zero, Selection, bias, post-encroachment, time

Researchers Geeta Singh and Neelu Khare of Vellore Institute of Technology in India propose an Entropy-Guided Jaya (EGJ) framework for feature selection in DDoS detection tasks. Their work, published in Frontiers in Communications and Networks, addresses a persistent weakness in wrapper-based feature selection: most metaheuristic search algorithms begin with randomly initialized candidate solutions, ignoring prior knowledge about which features actually matter. This random start can slow convergence and increase the risk of the search settling on mediocre feature subsets before it finds the informative ones.

The core innovation is a Weighted Composite Entropy (WCE) measure that fuses two complementary information-theoretic criteria. Shannon entropy quantifies the intrinsic variability and information content of a feature, while mutual information measures how much knowing that feature reduces uncertainty about the class label. Neither alone tells the full story: a feature may be highly variable yet useless for discrimination, or strongly predictive but information-poor. By combining them with tunable weights, WCE prioritizes features that are both information-rich and highly discriminative, producing a more balanced estimate of feature importance than either metric in isolation.

These entropy scores then seed the Jaya algorithm, a parameter-free population-based optimizer that iteratively moves each candidate solution toward the best solution in the population while pushing it away from the worst. Because feature selection is inherently binary, each feature is either kept or discarded, the researchers employ a binarized variant of Jaya enhanced with a Swish-based transfer function. The Swish function, smooth and non-monotonic, converts continuous solution vectors into binary feature masks through stochastic thresholding, preserving population diversity better than conventional step-function binarization.

An adaptive mutation strategy further guards against premature convergence. Mutation probability is adjusted according to the fitness gap between each candidate and the current best solution, allowing poorly performing individuals to explore the search space more aggressively while protecting high-quality subsets. Periodic random replacement every three generations injects fresh search directions, and elite solutions are retained so the best feature subsets are never lost. Checkpointing and early stopping round out the machinery, enabling recovery of long-running experiments and termination once no further improvement occurs.

To evaluate candidate feature subsets, the framework employs XGBoost, a gradient-boosted ensemble classifier implemented on the H2O distributed platform. The choice is deliberate: XGBoost offers robustness and scalability on large security datasets, and H2O’s distributed architecture reduces hardware requirements while supporting parallel model training. Fitness is assessed using macro recall across three-fold cross-validation, a metric that ensures balanced optimization across all attack categories, including minority classes often underrepresented in multiclass intrusion detection data.

The experimental benchmark is CIC-DDoS2019, a widely adopted dataset from the Canadian Institute for Cybersecurity containing realistic traffic traces across multiple reflection, amplification, and flooding attacks. After preprocessing, which removed non-informative attributes, duplicate records, invalid values, and 30 highly correlated features, the dataset retained 47 features and 259,641 balanced samples spanning eight attack types plus benign traffic. Running EGJ on this preprocessed data selected just 27 features, a substantial reduction that maintained high precision, recall, and F1-score while achieving an average accuracy of 93 percent.

Particularly notable is the framework’s performance on SSDP and UDP flood attacks, two categories that are notoriously difficult to distinguish because both rely on UDP communication and exhibit similar flow-level characteristics such as short durations, high packet rates, and bursty traffic patterns. EGJ achieved 76 percent precision and 71 percent recall for SSDP, and 75 percent precision with 77 percent recall for UDP, outperforming the baseline Multi-Island Jaya algorithm and several distributed detection frameworks. Among the selected features, attributes like DestinationPort, FlowBytes/s, PacketLengthStd, and TotalBackwardPackets capture the packet burstiness, flow asymmetry, and temporal behavior that separate amplified reflection traffic from direct flooding.

Statistical validation confirms the gains are real rather than artifacts of random variation. A one-sided exact Wilcoxon signed-rank test on paired per-class metrics yielded p-values of 0.03125 for precision and 0.015625 for recall and F1-score, all below the 0.05 significance level. The improvement of roughly 3 percent over the baseline carries substantial operational weight: when monitoring millions of network flows, even a small increase in detection accuracy translates to thousands of additional malicious flows correctly identified. Complexity analysis shows EGJ introduces only linear overhead, preserving the same asymptotic time and space requirements as conventional Jaya variants while improving convergence behavior.

Cross-dataset evaluation on UNSW-NB15, a more diverse multi-class intrusion benchmark, yielded 83.38 percent accuracy without modifying the optimization strategy, suggesting the entropy-guided approach generalizes beyond DDoS-specific traffic. The authors acknowledge that a comprehensive ablation study isolating the individual contribution of each component, Swish binarization, adaptive mutation, entropy-guided initialization, and the WCE formulation, remains future work, as does validation on additional benchmarks like CICIoT2023 and integration with online streaming learning frameworks for real-time distributed detection environments.

The broader significance lies in demonstrating that information-theoretic feature evaluation need not remain a standalone filter step but can be woven directly into wrapper-based metaheuristic search. By biasing the initial population toward informative features, EGJ starts the optimization from higher-quality candidate solutions, converges faster, and identifies compact yet discriminative feature subsets. For security teams wrestling with the scale and sophistication of contemporary DDoS attacks, from DNS and NTP amplification to SYN floods and application-layer WebDDoS, the result is a lightweight yet effective detection pipeline that reduces dimensionality, cuts computational cost, and sharpens the boundary between benign traffic and attack, even where attack categories overlap most stubbornly.

Subject of Research: Entropy-Guided Jaya optimization for robust feature selection in high-dimensional network attack detection

Article Title: Entropy-Guided Jaya optimization for robust feature selection in high-dimensional network attack detection

Article References: Singh, G., & Khare, N. (2026). Entropy-Guided Jaya optimization for robust feature selection in high-dimensional network attack detection. Frontiers in Communications and Networks, 7, Article 1886520. https://doi.org/10.3389/frcmn.2026.1886520

Image Credits: AI Generated

DOI: 10.3389/frcmn.2026.1886520

Keywords: Entropy-Guided, Jaya, optimization, robust, feature, selection, high-dimensional, network, attack, detection, scientific research

The research team, led by endocrinologists Xueli Ren and Li Quan, combed through the electronic health records of 2,634 patients exposed to immune checkpoint inhibitors between January 2020 and July 2025. From this pool, 812 patients had both a normal thyroid-stimulating hormone (TSH) level and a normal free thyroxine (FT4) measurement in the 90 days before their first infusion, making them clean starting points for tracking thyroid-axis changes. After applying the remaining baseline criteria, 627 patients had at least 365 days of administrative follow-up opportunity. But here is where the study takes an unusual and technically important turn: only 240 of those 627 patients, or 38.3 percent, actually had a qualifying TSH retest within the first year after treatment.

This monitoring gap is the analytical heart of the paper. Among the 240 patients who did receive a retest, 30, or 12.5 percent, had a clinically adjudicated thyroid event within the study’s definitions, with a Wilson 95 percent confidence interval spanning 8.9 to 17.3 percent. That figure sounds like a one-year incidence rate, but the authors are emphatic that it is not. It is an observed detection proportion under incomplete monitoring. The distinction matters enormously for anyone trying to compare real-world studies of immune-related endocrine toxicity, because different centers test different patients at different times, and the resulting numbers are not directly comparable.

The depth of the monitoring shortfall becomes vivid when the researchers examine what happened to the 210 retested patients who had no recorded event. Of these, 196, or 93.3 percent, had their last available thyroid test before day 365. In other words, even among the monitored minority, most patients stopped being checked well before the end of the follow-up year. Laboratory surveillance in routine clinical practice is driven by symptoms, ordering habits, and competing clinical priorities, not by a standardized protocol. Any thyroid dysfunction that emerges after the last test simply goes unrecorded, invisible to retrospective analyses built on electronic health records.

So who was more likely to be retested, and who was more likely to have an event recorded? The team ran exploratory logistic regressions with Benjamini-Hochberg adjustment for multiple comparisons, a statistical technique that controls the false discovery rate when many baseline variables are tested simultaneously. Two associations stood out after adjustment. Patients with thyroid antibody positivity in the baseline window had dramatically higher odds of a recorded event, with an odds ratio of 8.73 and a 95 percent confidence interval of 3.44 to 22.14, yielding a Benjamini-Hochberg q value below 0.001. Higher baseline TSH also predicted recorded events, with an odds ratio of 1.69 per 1 mIU/L and a confidence interval of 1.25 to 2.27, with a q value of 0.0027. Both associations persisted in a limited-adjustment Firth model, a penalized likelihood approach designed for situations where events are sparse and conventional maximum likelihood estimation can fail or produce unstable estimates.

The antibody finding has a plausible biological mechanism. Thyroid peroxidase and thyroglobulin antibodies signal that the immune system has already broken tolerance to thyroid antigens. When PD-1 inhibitors release the molecular brakes on T cells, this pre-existing autoreactivity can flare into destructive thyroiditis, often producing a transient hyperthyroid phase followed by permanent hypothyroidism requiring lifelong levothyroxine replacement. The baseline TSH association is more subtle. A TSH level at the upper edge of normal may reflect early subclinical thyroid failure, so patients with higher baseline values may be closer to the threshold at which dysfunction becomes detectable. But the authors caution that these exploratory associations may partly reflect differential detection rather than pure causation, because patients with known risk factors may simply be monitored more intensively, making events more likely to be found.

To address this selection problem, the researchers deployed inverse probability weighting, a statistical method that reweights the monitored patients so that their measured baseline characteristics mirror those of the entire baseline-eligible cohort. The idea is to estimate what the recorded-event proportion would look like if the monitored group were representative of everyone. With capped weights to prevent a few extreme observations from dominating, the standardized recorded-event proportion came out at 12.6 percent, with a bootstrap 95 percent confidence interval of 8.5 to 17.1 percent. The near-identity of this standardized figure with the raw 12.5 percent suggests that, at least for the measured baseline variables, the monitored patients were not dramatically different from the full cohort. But the method can only adjust for variables that were actually measured, and unmeasured drivers of monitoring selection remain a residual concern.

Sensitivity analyses probed the robustness of the findings from different angles. When the team required paired retesting, meaning both TSH and FT4 had to be rechecked rather than TSH alone, the analysis recorded 30 events among 239 patients. When the baseline window was restricted to laboratory reports strictly before the calendar date of the first ICI infusion, excluding same-day results that might reflect treatment-day perturbations, the analysis recorded 22 events among 171 patients. The event counts shifted modestly across these stricter definitions, illustrating how sensitive real-world event tallies are to explicit outcome definitions and laboratory timing, a point the authors elevate to a central recommendation of the paper.

The study’s most consequential message may be methodological rather than clinical. The authors argue that real-world studies of thyroid toxicity after checkpoint inhibitor therapy should routinely report monitoring coverage, laboratory timing, and explicit outcome definitions. Without these elements, the reported proportions from different centers are essentially incommensurable, and pooled estimates across studies can be badly misleading. The team points to established reporting frameworks, including the STROBE guidelines for observational studies and the RECORD framework for studies using routinely collected health data, as templates for making monitoring selection transparent. A center that tests every patient quarterly will report far higher event proportions than one that tests only symptomatic patients, even if the underlying biology is identical.

For clinicians managing patients on PD-1 inhibitors, the practical implications are nonetheless tangible. Roughly one in eight monitored patients developed a study-defined thyroid event within a year, and the strongest predictors were pre-existing thyroid autoantibodies and a baseline TSH at the higher end of normal. Patients with these features may warrant particularly vigilant laboratory surveillance, since early detection of hypothyroidism allows prompt levothyroxine initiation and prevents the fatigue, weight gain, and cold intolerance that can erode quality of life during cancer treatment. Conversely, the study also warns against overinterpreting event proportions from centers with incomplete monitoring, because the true burden of thyroid dysfunction in the unmonitored majority remains unknown. As checkpoint inhibitors continue to expand across tumor types, the study stands as a reminder that the thyroid axis is a frequent casualty of immunotherapy, and that honest accounting of endocrine toxicity requires confronting not just the biology of immune attack, but the messy realities of who gets tested and when.

Subject of Research: Thyroid dysfunction after PD-1 inhibitor therapy: selection into post-treatment monitoring and thyroid-axis trajectories in a single-centre real-world cohort study

Article Title: Thyroid dysfunction after PD-1 inhibitor therapy: selection into post-treatment monitoring and thyroid-axis trajectories in a single-centre real-world cohort study

Article References: Ren, X., & Quan, L. (2026). Thyroid dysfunction after PD-1 inhibitor therapy: selection into post-treatment monitoring and thyroid-axis trajectories in a single-centre real-world cohort study. BMC Endocrine Disorders. https://doi.org/10.1186/s12902-026-02650-y

Image Credits: AI Generated

DOI: 10.1186/s12902-026-02650-y

Keywords: Thyroid, dysfunction, PD-1, inhibitor, therapy, selection, post-treatment, monitoring, thyroid-axis, trajectories, single-centre, real-world