The research, led by Amgad Droby of Tel Aviv Sourasky Medical Center together with colleagues at the Feinstein Institutes for Medical Research in New York and collaborators at several other institutions, focused on a question that has haunted the field of neurodegeneration for decades: can we see Parkinson’s disease before it becomes Parkinson’s disease? The answer, according to this work, appears to be a qualified but encouraging yes. The team analyzed 295 resting-state functional MRI datasets drawn from 52 people with early-stage Parkinson’s disease, 92 first-degree relatives who carry Parkinson’s-associated mutations in the GBA1 or LRRK2 genes but show no symptoms, and 58 healthy non-carriers who served as controls.

The technical heart of the study lies in how the researchers processed those scans. Resting-state functional MRI measures spontaneous fluctuations in blood oxygenation across the brain while a person lies still, doing nothing in particular. These fluctuations reveal the brain’s intrinsic functional architecture: the networks of regions that rise and fall in activity together. Rather than examining individual connections one by one, the team used independent component analysis, a mathematical technique that decomposes the massive four-dimensional imaging data into spatial patterns representing distinct functional networks. This approach allows researchers to capture whole-network topographies, the characteristic spatial layouts of brain activity, rather than isolated pairwise links.

What makes this study particularly clever is its borrowing from a different imaging modality. For years, researchers studying Parkinson’s have relied on positron emission tomography with the fluorodeoxyglucose tracer to identify two reproducible metabolic patterns: the Parkinson’s disease-related pattern, known as PDRP, which is elevated in the motor phase of the illness, and the cognition-related pattern, or PDCP, which tracks cognitive decline. These patterns have proven valuable for diagnosis and for measuring disease progression in clinical trials, but PET imaging involves radiation exposure and is expensive and logistically demanding. The central question of the new work was whether equivalent patterns could be extracted from radiation-free MRI scans, and whether they would behave the same way in people who do not yet have symptoms.

To find out, the researchers combined the independent component analysis with a support vector machine, a classical machine learning classifier trained to distinguish the brain patterns of Parkinson’s patients from those of healthy non-carriers. The classifier achieved 87 percent accuracy in that primary discrimination task, with an average accuracy of 78 percent across the full set of group comparisons. Once the classifier had learned to separate the groups, the team applied local interpretable model-agnostic explanations, a technique designed to open the black box of machine learning and reveal which features of the data drove each decision. This step isolated two fMRI-derived patterns that the investigators named fPDRP and fPDCP, the functional analogues of the well-known PET patterns.

With these patterns in hand, the researchers computed expression scores for every participant, essentially asking how strongly each individual’s brain expressed the Parkinson’s-related and cognition-related network topographies. The results were striking. Patients with early Parkinson’s disease showed significantly higher expression of both fPDRP and fPDCP than any of the preclinical groups. But within the preclinical population, an important divergence emerged. Participants classified as having a high likelihood risk, defined as a score of 60 or above on the Movement Disorder Society’s criteria for prodromal Parkinson’s, showed elevated fPDRP expression compared with lower-risk carriers, while their fPDCP expression remained unchanged. In other words, the motor-related network signature appears to rise first, before the cognitive signature, mirroring the clinical course of the disease itself.

Perhaps the most clinically consequential finding concerns what did not matter. Neither the specific genetic mutation a carrier carried, whether GBA1 or LRRK2, nor the status of their cerebrospinal fluid alpha-synuclein seed amplification assay, a molecular test that detects misfolded alpha-synuclein protein, significantly affected network expression in people who already had clinical Parkinson’s disease. This suggests that once the disease manifests, the functional network disruption converges on a common topography regardless of the underlying biology that triggered it. The network patterns, in this sense, capture the final common pathway of the disease rather than its upstream causes.

The team did not stop at cross-sectional analysis. To test whether the fMRI-derived patterns were stable and meaningful over time, they validated their findings in an independent cohort of 33 non-manifesting carriers who had been followed for ten years as part of a long-term observational study. The patterns proved reproducible in this validation cohort, lending weight to the idea that they represent genuine biological signals rather than statistical artifacts. Among the validation participants, eight individuals converted to manifest Parkinson’s disease during the follow-up period. Those converters showed numerically higher baseline fPDRP expression than those who did not convert, though the difference did not reach statistical significance, likely because of the small number of converters. The direction of the effect, however, is exactly what one would hope to see in a genuine preclinical marker.

The implications of this work extend well beyond the laboratory. Drug development for Parkinson’s disease has been repeatedly frustrated by the difficulty of testing therapies in people whose brains are already substantially damaged by the time of diagnosis. A reliable, radiation-free marker of preclinical network disruption could allow trials to enroll people at the earliest stages of the disease process, when interventions have the best chance of altering its trajectory. Because resting-state fMRI involves no ionizing radiation, it could in principle be repeated frequently, enabling researchers to track network changes over time in a way that PET-based measures cannot easily support. The study was supported by Biogen and by the Michael J. Fox Foundation for Parkinson’s Research, both of which have invested heavily in the search for reliable biomarkers of early disease.

Cautions remain, of course. The high-risk group in the preclinical analysis was defined by clinical criteria rather than by certain future diagnosis, and the conversion analysis, while directionally encouraging, involved only eight individuals. The authors themselves note that the fPDRP differences between converters and non-converters were not statistically significant. Still, the convergence of evidence across independent cohorts, multiple genetic risk groups, molecular CSF testing, and a decade of longitudinal follow-up makes a compelling case that the brain’s functional network architecture carries readable information about Parkinson’s disease risk long before the first tremor. For the millions of people who carry Parkinson’s-associated mutations and live with the uncertainty of what the future holds, the prospect of seeing the disease coming, and perhaps one day stopping it before it starts, has moved measurably closer.

Subject of Research: Resting-state fMRI detection of Parkinson's disease-related functional network patterns in preclinical and early disease stages

Article Title: Resting state fMRI network topographies in preclinical and early Parkinson’s disease stages

Article References: Droby, A., Nguyen, N., Do, P., Truong, J., Marebwa, B., Cedarbaum, J. M., Mirelman, A., Eidelberg, D., Vo, A., & Thaler, A. (2026). Resting state fMRI network topographies in preclinical and early Parkinson’s disease stages. npj Parkinson's Disease. https://doi.org/10.1038/s41531-026-01584-5

Image Credits: AI Generated

DOI: 10.1038/s41531-026-01584-5

Keywords: Parkinson's disease, resting-state fMRI, biomarkers, GBA1, LRRK2, machine learning, independent component analysis, PDRP, PDCP, alpha-synuclein, prodromal Parkinson's, neuroimaging

The site itself is dramatic. Two sets of paintings, designated C1 and C2, adorn the inner walls of naturally formed caves along an east-west oriented granite ridge in Fuyun County, roughly 70 kilometres from the county town. The two panels are separated by about 40 metres of vertical elevation and 112 metres of horizontal distance, and a circular stone mound tomb built of white pebbles lies approximately 100 metres from the cave entrances. Panel C1 stretches roughly 20 metres in length and 11.8 metres in width, while C2 contains 51 independent compositions distributed along both walls and even onto the ceiling. The paintings’ good state of preservation, unusual for open rock art in this harsh environment, is precisely what makes the site such a valuable archive.

The technical heart of the study lies in how the team saw the invisible. Working in two stages, the researchers first captured high-resolution digital photographs of the entire rock art panels using a Canon EOS 5D Mark IV camera, magnifying the images block by block to flag suspicious micro-traces and areas of superimposition. These regions of interest were then scanned with a Specim IQ hyperspectral camera, which records reflected light in 180 narrow spectral bands between 400 and 930 nanometres, far beyond the discrimination of human vision. After radiometric calibration against a standard white reference and removal of noisy bands, the resulting data cubes were processed using Independent Component Analysis, a statistical method that performs blind source separation to extract distinct pigment signatures, and the Spectral Angle Mapper algorithm, which classifies regions sharing identical or similar spectral compositions using a maximum spectral angle threshold of 0.8 radians.

The payoff was immediate and striking. In panel C2, hyperspectral imaging revealed two motifs, labelled C2-50 and C2-51, that are completely invisible to the naked eye. Crucially, one visible figure, C2-41, was found to be superimposed over the hidden C2-51. The ICA components showed that the red and green areas correspond to genuinely different pigments rather than variations of a single material, and the spectral curves confirmed the distinction. Because the outlines of C2-41 are continuous and intact while C2-51 appears fragmented and truncated, the team concluded that the hidden motifs predate the horned anthropomorphic figures that dominate the panel, evidence of a painting sequence older than anything previously documented at the site.

The same analytical machinery resolved a long-standing ambiguity in panel C1. One blurred motif, C1-12, had resisted identification for years, its faint traces lost against the naturally reddish granite. Hyperspectral processing revealed it to be a crossed vajra, one of the most recognisable symbols in the Tibetan Buddhist visual system, representing indestructible wisdom and the firmness of the dharmadhatu. Alongside it, the panel preserves three instances of Sanskrit mantras rendered in Tibetan script, complete with termination marks and, in one case, a Tibetan seed syllable, features central to the textual traditions of Buddhist mantra practice. These are not casual doodles but technically literate reproductions of an institutionalised religious vocabulary.

Dating the imagery required a different toolkit. In the absence of direct scientific dating, the researchers built a relative chronology through iconographic comparison anchored to two benchmarks: the Chemurchek culture, which flourished between roughly 2500 and 1700 BCE, and the broader Early Bronze Age from about 3000 to 1000 BCE. The bell-shaped anthropomorphic figure C1-2, positioned at the centre of C1’s rear wall, closely resembles figures carved on stone slabs at the Khar Chuluut-1 monumental ceremonial enclosure in Mongolia’s Bayan-Ölgii Province, structures that radiocarbon dating places before the mid-third millennium BCE. A comparable motif also appears at Dundebulake-1 in Altai Prefecture, Xinjiang. Because northern Xinjiang formed part of the wider Chemurchek cultural interaction sphere, the team attributes C1-2 to the early Chemurchek period, making it one of the oldest images in the cave.

The horned anthropomorphic figures that crowd panel C2 tell a slightly later story. These figures, marked by short horn-like lines on the head, frequently cluster around X-shaped symbols and appear in groups; some display rounded abdomens with small circular marks that earlier scholars have interpreted as possible indications of pregnancy, while three figures hold bows and arrows. Their closest parallels lie in the Samus, Karakol, and Okunev cultures of the Altai and southern Siberian Bronze Age, spanning roughly 2500 to 1000 BCE. The researchers also demonstrated that apparent colour differences among some figures, such as motifs C2-45 through C2-48, are deceptive: ICA analysis confirmed they share the same pigment composition, with the visual variation caused by water rusting rather than separate painting episodes, a cautionary lesson in how weathering can manufacture false chronologies.

Perhaps the most tantalising question concerns the three-eyed face C1-1, with its seventeen radiating head lines, concentric-circle eyes, a third eye on the forehead, and an arc-shaped beard. Its formal traits echo the sun-headed and multi-eyed figures of the Okunev tradition, often discussed within shamanic interpretive frameworks, and some team members noted a visual resemblance to Tibetan Buddhist wrathful deities such as White Mahakala. Yet the authors are deliberately cautious. The weathering of C1-1 is substantially heavier than that of the confirmed Buddhist motifs, and its pigment appearance differs markedly, arguing for a different and earlier phase of execution. The similarity, they conclude, is best understood as formal resemblance rather than evidence of religious continuity, and the motif’s precise meaning and dating remain open pending scientific dating.

The Tibetan Buddhist layer, by contrast, can be situated historically. Under the Yuan dynasty, which ruled from 1271 to 1368, imperial patronage carried Tibetan Buddhist ideas and iconography into frontier zones well beyond the Tibetan Plateau, and a second wave of transmission accompanied the eastward migration of the Torghut tribe during the mid to late Qing dynasty, which brought lamas and monasteries into the Altai. Within this framework, the researchers hypothesise that the Buddhist motifs at Tangbaletasi could have first appeared as early as the Yuan period. Notably, the later painters did not erase what came before: earlier images remained visible beneath and beside the mantras and vajra-crosses, producing a genuine coexistence of visual systems within a single cave space.

What was the cave for? The team is careful to frame its answer as a working hypothesis grounded in converging spatial evidence rather than a single smoking gun. The site conforms to regional patterns documented for anthropomorphic rock art across the Altai, occupying a consistent elevation band, lying close to water sources, and opening onto a south-facing rock face. Its proximity to a burial mound echoes wider Inner Asian associations between image-bearing sites and commemorative landscapes. Inside, the compositions, horned figures arranged in clusters around fire-like symbols, symbolic avatars rather than portraits of real people, suggest collective, non-domestic activity rather than everyday life. The authors acknowledge their limitations honestly: no direct scientific dating was performed, and no excavation data underpin the functional interpretation. But the picture that emerges is compelling, a mountain shelter where Bronze Age ritual specialists, and centuries later Buddhist practitioners, each found reason to leave their mark on the same sacred stone, layer upon layer, across four thousand years of human belief.

Subject of Research: Chronology and function of the Tangbaletasi Cave Paintings in Xinjiang's Altai region using hyperspectral imaging and iconographic analysis

Article Title: Chronology and function of Tangbaletasi cave paintings via hyperspectral and iconographic analyses

Article References: Huang, F., Fu, Y., Zhou, G., Li, L., Chai, Y., Liu, C., Wang, J., Peng, J., & Zhuoya, B. (2026). Chronology and function of Tangbaletasi cave paintings via hyperspectral and iconographic analyses. npj Heritage Science, 14(1), Article 540. https://doi.org/10.1038/s40494-026-02433-7

Image Credits: AI Generated

DOI: 10.1038/s40494-026-02433-7

Keywords: Tangbaletasi Cave Paintings, rock art, hyperspectral imaging, Altai Mountains, Bronze Age, Chemurchek culture, Tibetan Buddhism, Yuan dynasty, Independent Component Analysis, pigment analysis, archaeology, Xinjiang

The study used the Upper Limb Movement EEG Dataset, recorded at Graz University of Technology, which the authors argue is uniquely suited to this kind of controlled investigation. The dataset contains recordings from fifteen healthy subjects, aged 22 to 40, who performed six distinct movements of the right upper limb, including elbow flexion and extension, forearm supination and pronation, and hand opening and closing, with the arm supported by an exoskeleton to prevent muscle fatigue. Crucially, the data were collected with sixty-one active wet electrodes at a high sampling rate of 512 Hz and distributed in a completely raw format. That raw provenance matters enormously here: many widely used EEG benchmarks arrive with hidden preprocessing already applied, which makes it impossible to disentangle what the model is actually learning from what the pipeline has already cleaned up for it.

The experimental design was deliberately ablative. The researchers trained four architectures, the convolutional EEGNet alongside three Transformer models designed for EEG work: ContraNet, built for motor imagery decoding; Conformer, designed for motor imagery and emotion decoding; and EEGDeformer, formulated for cognitive attention detection. Each model was tested on binary classification, distinguishing movement from rest, and on a harder three-class problem separating hand opening, hand closing, and rest, under four combinations of preprocessing and normalization. Preprocessing involved a rigorous pipeline: the PREP procedure to remove and interpolate noisy channels, robust common average referencing, band-pass filtering between 1 and 100 Hz, notch filtering of power-line noise and its harmonics, and artifact removal through Independent Component Analysis using the Extended Infomax algorithm with automatic classification of artifactual components by ICLabel. No epochs were discarded, preserving perfect class balance.

The headline numbers are striking. In motor execution binary classification under the subject-mixed protocol, EEGDeformer improved from 0.51 accuracy without normalization and preprocessing to 0.89 with both applied, a relative gain of roughly 75 percent. Conformer and ContraNet showed similar trajectories, climbing from near-chance levels of around 0.52 to 0.87 and 0.85 respectively. Even EEGNet, the compact convolutional baseline, jumped from 0.55 to 0.86. Without normalization, models frequently failed to exceed the statistical significance threshold of 56 percent established for the binary task, meaning their apparent learning was indistinguishable from chance. The pattern repeated in motor imagery, though with smaller absolute gains, and in the three-class problems, where executed movements reached up to 0.66 accuracy with ContraNet while imagined movements proved considerably harder.

Why should a Transformer care so much about whether its inputs have zero mean and unit variance? The authors point to the mathematics of attention itself. Transformer models compute attention weights by passing scaled dot products of query and key vectors through a softmax function. EEG signals vary wildly in amplitude across subjects because of differences in skull thickness, electrode impedance, and neural activation strength. Feed such unstandardized signals into the attention mechanism and the dot products become extreme, saturating the softmax so that nearly all weight collapses onto a single token while the rest of the spatiotemporal sequence is effectively ignored. The model is then blind to the distributed patterns that characterize motor activity in the brain. Z-score normalization, computed channel by channel as the raw value minus the mean divided by the standard deviation, keeps those dot products in comparable ranges across subjects, allowing nuanced attention patterns to emerge.

Perhaps the most important finding, however, is that normalization alone is not enough. When Z-score scaling was applied to non-preprocessed, artifact-contaminated signals, it provided minimal benefit and occasionally made things worse. The explanation is proportional compression: when artifacts dominate the variance of a signal, the normalization parameters are determined primarily by noise rather than by neural patterns, effectively lowering the signal-to-noise ratio at the model’s input. Only after rigorous artifact removal through ICA and the PREP pipeline do the normalization statistics reflect genuine neural variability. Preprocessing and normalization thus form a synergistic pair, a two-stage sequence in which artifacts are removed first to improve signal quality, and the cleaned signals are then standardized across subjects while preserving their relative temporal and spatial structure. Neither step alone suffices; together they enabled every model to reach statistically significant performance.

The study also distinguished between two validation regimes with important practical implications. In the subject-mixed protocol, data from all participants were pooled and split, allowing intra-subject information to leak across subsets and representing a theoretical upper bound. In the stricter subject-independent protocol, twenty percent of participants were held out as a test group, and normalization parameters were fitted exclusively on training subjects and applied to validation folds, preventing any cross-subject leakage. For final evaluation on unseen test subjects, Z-score parameters were derived from the test subjects’ own unlabeled data, simulating the standard unsupervised calibration phase a new BCI user would undergo. Encouragingly, the improvements held under the strict protocol, with EEGDeformer reaching 0.85 in motor execution binary classification and all models surpassing their significance thresholds, confirming that normalization facilitates learning of genuinely generalizable, subject-independent features.

Training dynamics told a consistent story. Models trained with normalized data converged faster, with training loss dropping more steeply in early epochs, and their validation loss curves showed less oscillation and a smaller gap between training and validation performance, indicating reduced overfitting. Architecture mattered too: EEGDeformer, with its dense connections that propagate input scaling effects throughout the network, showed the largest relative gains, while ContraNet’s hybrid convolutional-transformer design delivered the best absolute multiclass motor execution accuracy. Notably, the benefits extended beyond attention-based models, with the convolutional EEGNet improving by 51 percent relative in motor execution binary classification, suggesting that input standardization aids learning across architecture families, though Transformers appear especially sensitive to it because their learned positional encodings can be disrupted by unstandardized inputs.

The authors are candid about limitations. The analysis relied on a single dataset, albeit one chosen precisely because its raw format enables a rigorous ablative study, and only a binary choice of applying or not applying Z-score was tested, leaving alternatives such as min-max scaling, robust scaling, or frequency-band-specific normalization unexplored. The normalization strategies also assume either offline processing or an unsupervised calibration buffer, and fully online pipelines with adaptive Z-score calibration remain future work. Still, the central conclusion stands with unusual force: Z-score normalization is not an optional technical detail but a critical requirement, on par with architecture selection itself, for Transformer-based EEG motor classification. For a field racing toward practical brain-controlled prosthetics and assistive devices, the message is that the humblest step in the pipeline may deserve the most attention.

Subject of Research: Effects of Z-score data normalization on Transformer-based EEG motor imagery and motor execution classification

Article Title: Effects of EEG-data normalization on EEG-Transformer-based motor classification

Article References: Mattei, E., & Lozzi, D. (2026). Effects of EEG-data normalization on EEG-Transformer-based motor classification. Multimedia Tools and Applications, 85(10), Article 796. https://doi.org/10.1007/s11042-026-21929-9

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21929-9

Keywords: EEG, brain-computer interface, Transformer, Z-score normalization, motor imagery, motor execution, deep learning, preprocessing, independent component analysis, subject-independent validation, EEGDeformer, Conformer

The conceptual leap behind CSI-TME borrows from genetics. Just as synthetic lethality describes pairs of genes whose simultaneous loss is lethal even when either gene alone is dispensable, the researchers sought pairs of transcriptional states from different cell types whose joint activity, rather than either state alone, is associated with clinical outcome. The classic example of a synthetic lethal pair is PARP1 and BRCA1/2, which underpins the use of PARP inhibitors in BRCA-mutant breast cancer. Importantly, the authors note that such pairs need not interact physically, and likewise a cell-state interaction may be mediated by mechanisms other than a direct ligand-receptor handshake. This framing allowed the team to hunt for clinically meaningful crosstalk in data that were never designed to capture it.

Technically, the pipeline proceeds in three stages. First, it uses a deconvolution tool called CODEFACS, guided by cell-type marker signatures derived from single-cell data, to split bulk expression profiles from each patient into cell-type-specific expression matrices. The team worked with 60,751 cells from IDH-mutant glioma samples to define seven cell types: malignant cells, T cells, B cells, myeloid cells, endothelial cells, oligodendrocytes, and stromal cells. Second, for each cell type it applies independent component analysis to the deconvolved data, extracting ten independent components per cell type, each representing a distinct transcriptional state or gene expression program. Third, it screens all pairs of components from different cell types using Cox proportional hazards regression, testing whether their joint activity, binned as low-low, high-high, or high-low combinations, is associated with overall survival while controlling for the individual activities of each state and for demographic covariates such as age and sex.

Applied to 425 IDH-mutant glioma samples from The Cancer Genome Atlas, the method detected 160 significant cell-state interactions at a false discovery rate threshold of 20 percent, with 70 percent internal cross-validation accuracy. Strikingly, about 70 percent of these interactions were associated with worse survival, meaning the interaction network is predominantly pro-tumor. The findings were validated in an independent cohort of 325 IDH-mutant glioma patients from the Chinese Glioma Genome Atlas: interactions identified as pro-tumor in the discovery cohort tended to carry positive hazard ratios in the validation cohort, and 51 percent of the interactions could be independently rediscovered in CGGA without relying on the TCGA-derived factorization, compared with a median recovery of only 7 percent for randomized controls. Malignant cells and B cells participated in the greatest number of interactions, and most interactions involving T-cell states were pro-tumor, consistent with tumor-driven T-cell dysfunction.

Among the malignant cell states recovered by the analysis were programs resembling known glioma lineages, including astrocyte-like and oligodendrocyte-progenitor-like states. One component stood out: its negative signature genes overlapped two independent glioma stemness signatures, included the neurodevelopmental transcription factor SOX11 and the stem-cell-associated gene CD44, and were highly expressed during embryonic human brain development before declining in fetal and adult tissue. The team concluded that this component captures highly proliferative glioma stem cells. These stem-like malignant cells engaged in pro-tumor interactions with several immune states. One particularly intriguing pairing linked glioma stem cells with a T-cell state whose signature genes were enriched for proliferation and, unexpectedly, for senescence markers, suggesting that proliferating T cells in the tumor microenvironment may become senescent, while stem-like tumor cells maintain their proliferative capacity, a synergy associated with worse survival. Another interaction suggested that when glioma stem cells are reduced and T-cell interferon signaling is simultaneously dampened, prognosis worsens, with gene-level analysis implicating interferon genes such as IFIT1 and IFIT3 in T cells alongside the immune-regulatory gene PTN in malignant cells.

The two major subtypes of IDH-mutant glioma, astrocytoma and oligodendroglioma, showed differential interaction activity. The researchers quantified how often each interaction was active across patients, a metric they call penetrance, and found that interactions dominant in astrocytomas were skewed toward anti-tumor effects, an enrichment that persisted after controlling for the younger age of astrocytoma patients. When patients in both subtypes were stratified by whether pro-tumor or anti-tumor interactions dominated in their tumors, those dominated by anti-tumor interactions survived significantly longer in both TCGA and CGGA. This suggests that measuring the balance of cell-state interactions refines prognosis beyond the standard molecular classification.

Because a cell-state interaction inferred from joint activity need not reflect physical communication, the team examined how many interactions could be backed by known ligand-receptor pairs from the CellChat database. Roughly 20 percent of the interactions, 32 of 160, involved complementary ligand-receptor pairs, comprising 69 unique pairs. More convincingly, when the researchers scored six publicly available spatial transcriptomics datasets of IDH-mutant glioma, the ligand-receptor-supported interactions were significantly enriched among spatially proximal cell-state pairs across all six slides, while interactions overall showed no such spatial organization. The standout example was a pro-tumorigenic interaction between hypoxic, partially epithelial-to-mesenchymal-transitioning malignant cells and tip-like endothelial cells, the angiogenesis-specialized endothelial state, supported by 25 distinct ligand-receptor pairs and co-localized in every spatial dataset examined. Another spatially supported, anti-tumor interaction involved JAG2 on T cells and NOTCH2 on malignant cells, consistent with a role for NOTCH-mediated adhesion in T-cell anti-cancer immunity.

The clinical relevance of the network extended to therapy. In a pre-treatment transcriptomic dataset from 29 glioma patients undergoing neoadjuvant anti-PD1 immune checkpoint blockade, the penetrance of pro-tumor interactions was significantly higher in non-responders than in responders, while anti-tumor interactions showed the converse pattern with marginal significance. Comparing paired primary and recurrent biopsies from the GLASS consortium revealed that pro-tumor interaction load and penetrance increased significantly at relapse, with three specific pro-tumor interactions, two involving stromal cells, significantly enriched in recurrent tumors. The method also generalized beyond brain cancer. Applied to TCGA breast cancer, melanoma, and head and neck cancer, CSI-TME again found predominantly pro-tumor networks that were largely cancer-type-specific. In breast cancer, pro-tumor interactions were more penetrant in pre-malignant lesions that later progressed and in patients resistant to trastuzumab; in melanoma, anti-tumor interactions were more penetrant in patients responding to anti-PD1 and BRAF-inhibitor therapies; and in head and neck cancer, pro-tumor interactions were more penetrant in patients who failed to respond to cetuximab.

Perhaps the most conceptually provocative finding emerged from integrating the interaction network with somatic mutation data. Anti-tumor interactions were strongly over-represented among the 91 interactions significantly associated with mutated genes, and this association was concentrated in early-stage, lower-grade tumors, with the balance shifting toward pro-tumor interactions in advanced grades. The authors interpret this as evidence that the tumor microenvironment mounts a homeostatic, tissue-protective response to oncogenic mutations early in tumorigenesis, a response that is gradually reprogrammed toward tumor promotion as the disease progresses. Together with the observation that the network stratifies immunotherapy response and prioritizes targetable ligand-receptor communication, these results position CSI-TME as a practical route to extracting single-cell-level insight from the vast archives of bulk clinical transcriptomic data that already exist, offering a new lens on how the ecosystem surrounding a tumor shapes its course. The pipeline is freely available to the research community.

Subject of Research: Computational inference of clinically relevant cell-state interactions in the tumor microenvironment of IDH-mutant gliomas

Article Title: Identifying clinically relevant cell state interactions in the tumor microenvironment of IDH-mutant gliomas using CSI-TME

Article References: Identifying clinically relevant cell state interactions in the tumor microenvironment of IDH-mutant gliomas using CSI-TME. (n.d.). https://doi.org/10.1038/s44320-026-00201-0

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00201-0

Keywords: IDH-mutant glioma, tumor microenvironment, cell-state interactions, CSI-TME, bulk RNA-seq deconvolution, independent component analysis, glioma stem cells, ligand-receptor signaling, immunotherapy response, spatial transcriptomics, synthetic lethality, tumor evolution

The core of the research is a framework the authors call ICO-2DQN, a hybrid of an Intelligent Cuckoo Optimization algorithm and a Dueling Deep Q-Network. The system was trained on 4,000 rows of biomechanical data recorded from martial artists performing fundamental techniques—chopping, kicking, and grappling—while wearing IoT sensors. Each record captured a single movement instance, with sensor-derived features describing joint motion, force distribution, body stability, and movement performance. The goal was ambitious: build a machine that could not merely classify a movement as good or bad, but generate individualized corrective feedback the way an expert coach would, only faster and with quantitative precision.

Before any learning could happen, the raw sensor data required careful preparation. The researchers handled missing values to prevent biased biomechanical estimates, then applied min–max normalization to scale every heterogeneous feature onto a common range from zero to one. This step matters more than it might appear: joint angles measured in degrees, forces measured in newtons, and velocities measured in radians per second live on wildly different scales, and without normalization a single large-magnitude feature could dominate the learning process. The preprocessing ensured that every biomechanical signal contributed equally to the model’s understanding of a technique.

The next stage used Independent Component Analysis, a statistical technique that decomposes complex, mixed sensor signals into their underlying independent sources. Wearable sensors pick up a tangle of overlapping information—muscle-driven forces, gravitational effects, sensor noise, and the compound motion of linked body segments all arrive blended together. ICA separates that mixture, isolating distinct biomechanical patterns that would otherwise remain hidden. The researchers retained twelve independent components, a number chosen to reduce redundancy while preserving the essential structure of each movement. Compared with alternatives such as kernel ICA or autoencoders, linear ICA offered faster execution and lower computational cost, an important consideration for a system intended to run alongside real-time IoT coaching.

From the cleaned and decomposed signals, the system derived the biomechanical quantities that define martial arts technique: joint angles, angular velocities, center-of-mass stability, and force delivery. These four quantities form the state space of the reinforcement learning environment. At every time step, the learning agent observes a vector containing the practitioner’s joint angles, angular velocities, balance stability, and force distribution. Its action space consists of the corrective instructions a coach might issue—posture correction, balance adjustment, movement speed modification, or simply continuing the current motion. A weighted reward function scores each executed movement based on posture accuracy, balance stability, and movement similarity, so the agent learns to issue feedback that maximizes the quality of the student’s technique.

The Dueling Deep Q-Network at the heart of the system introduces a clever architectural refinement over standard deep Q-learning. Rather than estimating the value of every state-action pair directly, the dueling architecture decomposes the action-value function into two separate streams: a state-value function that captures how good a biomechanical position is in itself, and an advantage function that captures how much a particular action improves on the average. This decoupling is well suited to martial arts, where many states have similar values regardless of the action taken, and it accelerates convergence in the high-dimensional, continuous movement spaces that full-body biomechanics produce. The network uses convolutional layers for feature extraction before splitting into the two value streams, which are then recombined into Q-values that guide the selection of optimal training strategies.

Reinforcement learning systems, however, are notoriously sensitive to their hyperparameters—learning rates, discount factors, and the weights inside the reward function. Poorly tuned, an agent converges slowly or oscillates between policies. That is where the cuckoo enters. The Intelligent Cuckoo Optimization algorithm draws its inspiration from the brood parasitism of real cuckoos, which lay their eggs in the nests of host birds. In the algorithmic version, candidate solutions are cuckoo eggs, each habitat receives a variable number of eggs within defined bounds, and an egg-laying radius determines how far solutions scatter in the search space. The algorithm also incorporates Lévy flight behavior, the characteristic pattern of long jumps interspersed with short steps that cuckoos use when seeking new habitats. Applied here, the ICO dynamically tunes the 2DQN’s hyperparameters and reward weightings, replacing weak solutions with fitter ones and steering the learning process away from local optima.

The experimental results are striking. Implemented in Python and evaluated with five-fold cross-validation on the IoT sensor dataset, the ICO-2DQN model achieved an accuracy of 98.7 percent, a precision of 98 percent, a recall of 96.5 percent, and an F1-score of 97.1 percent, with a root-mean-square error of 0.15. In real-time testing, the system recorded an accuracy of 98.2 percent, an average F1-score of 0.981, and a feedback delay of just 82 milliseconds—fast enough for a student to receive a correction while the movement is still fresh. Statistical validation with paired t-tests and 95 percent confidence intervals confirmed the gains were significant at p < 0.001. During training, the model’s cumulative reward climbed from roughly 20 to 98 while training loss fell from about 1.12 to 0.04, and the exploration rate decayed from 1.0 to 0.1 as the agent shifted from exploration to exploitation.

The comparison against baseline models reinforces the case. When retrained and tested on the same dataset under identical conditions, competing frameworks—including two-stream CNNs, a 3D-CNN combined with LSTM, a Vision Transformer paired with a DQN, a Sunflower Optimization multi-column CNN, and classical support vector machines—fell short on accuracy, error rates, feedback latency, and stability. Each baseline carried known weaknesses: optical-flow dependence and computational cost in two-stream CNNs, limited interpretability in 3D-CNN plus LSTM models, heavy data and compute demands in ViT-DQN, and motion-blur sensitivity in YOLO plus LSTM approaches. The dueling architecture’s explicit modeling of sequential action-state transitions, such as the shift from a kick to a grappling exchange, proved better suited to the spatiotemporal dependencies of martial arts than any of the alternatives.

Evaluation across skill levels showed the framework adapting its behavior for beginner, intermediate, and advanced practitioners while maintaining high movement recognition accuracy and reliable biomechanical assessment throughout. The authors are candid about the remaining hurdles: the system depends on accurate, fine-grained motion data, and inconsistent or faulty sensor readings could corrupt the analysis; real-time deployment also demands robust wearable hardware that resource-limited training environments may lack. Still, the trajectory is clear. The researchers envision future systems that adapt to specific martial arts styles, body types, and skill levels, integrating live motion capture to adjust training programs automatically with an emphasis on both performance and injury prevention. Biomechanical monitoring can already flag irregular loading patterns and fatigue-related deviations before they become injuries, and a reinforcement learning agent that prescribes corrective exercises or modulates training intensity in advance could make practice substantially safer. If the approach matures, the centuries-old model of learning by imitation may gain a tireless digital partner—one that sees what no coach can, and answers in milliseconds.

Subject of Research: Deep reinforcement learning combined with biomechanical movement analysis for adaptive martial arts training

Article Title: Innovative martial arts teaching methods based on deep reinforcement learning and biomechanics of movement

Article References: Zheng, L., & Zhang, J. (2026). Innovative martial arts teaching methods based on deep reinforcement learning and biomechanics of movement. Discover Artificial Intelligence, 6(1), Article 1307. https://doi.org/10.1007/s44163-026-01887-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01887-9

Keywords: martial arts, deep reinforcement learning, biomechanics, Dueling DQN, cuckoo optimization, IoT sensors, Independent Component Analysis, sports training, movement analysis, personalized coaching, injury prevention, machine learning

The question the researchers set out to address is deceptively simple: do different chemotherapy regimens leave different signatures on brain function? This matters because so-called chemotherapy-related cognitive impairment, often described by patients as brain fog, has mostly been studied as a single phenomenon, as if all cytotoxic drugs affected the brain in the same way. If different regimens produce measurably different network alterations, that could eventually help clinicians predict which patients are most at risk and open the door to regimen-specific protective strategies. The authors are careful, however, to frame their results as regimen-associated associations rather than proof of direct causation, because the two patient groups differed not only in their drugs but also in their cancer types, metastatic patterns, and specific treatment combinations.

To probe the brain’s wiring, the team enrolled 43 colorectal cancer patients who had completed two to three months of fluorouracil-based chemotherapy and 54 non-small cell lung cancer patients who had received two to three months of carboplatin-based chemotherapy. At the end of treatment, each participant underwent resting-state functional MRI, a technique that records spontaneous, low-frequency fluctuations in blood-oxygen signals while the subject simply lies still in the scanner. Because these fluctuations are synchronized across regions that work together, the timing patterns can be used to reconstruct the brain’s functional connectome, a map of which areas are talking to which.

The first analytical approach was graph theory, a branch of mathematics that treats the brain as a network of nodes and edges. Each region of a standardized brain parcellation becomes a node, and the statistical correlation between its activity and that of other regions defines the edges. From this network the researchers computed nodal metrics such as degree, a measure of how many strong connections a region maintains, and global efficiency, which captures how easily that region can exchange information with the rest of the network. When the two patient groups were compared, one result stood out after correction for multiple comparisons, the statistical safeguard that guards against false positives when thousands of brain regions are tested simultaneously.

That robust finding centered on the left precuneus, a hub tucked into the medial parietal cortex that plays a central role in self-referential thought, memory retrieval, and consciousness itself. Patients treated with carboplatin-based chemotherapy showed decreased nodal degree and decreased nodal global efficiency in the left precuneus compared with patients who had received fluorouracil-based treatment. In plain terms, this key integrative hub appeared less connected and less efficient at relaying information in the carboplatin group. Because this result survived multiple comparison correction, the authors consider it the most reliable signal in the study, and it points to carboplatin-based regimens being associated with more pronounced topological alterations in this region.

The second line of analysis used group independent component analysis, a data-driven method that decomposes whole-brain signals into spatially coherent networks without imposing a prior template. This procedure identified the canonical resting-state networks familiar to systems neuroscientists, including the default mode network in its dorsal and ventral subdivisions, the salience network, the left and right executive control networks, the sensorimotor network, the language network, the auditory network, and the primary visual network. The researchers then examined how strongly these networks communicated with one another, both in a static sense, averaged across the entire scan, and dynamically, by tracking how connectivity patterns shift from moment to moment across short sliding windows.

The static connectivity comparisons revealed a pattern of reduced communication in the colorectal cancer group. Patients who had received fluorouracil-based chemotherapy showed decreased connectivity within the right cuneus, part of the precuneus network, and reduced inter-network connectivity between the dorsal default mode network and the salience network, between the dorsal default mode network and the language network, between the ventral default mode network and the primary visual network, and between the sensorimotor network and the primary visual network. The default mode network, which is active when the mind wanders and reflects, and the salience network, which decides which stimuli deserve attention, are both repeatedly implicated in cognitive complaints after cancer treatment, so a weakening of the bridge between them is intriguing.

The dynamic analysis added a temporal dimension that static measures cannot capture. By clustering the sequence of connectivity states across time, the team identified four recurring brain states, each representing a distinct temporary configuration of network communication. Patients in the fluorouracil group spent less time in one particular configuration, showing decreased mean dwell time and a reduced fraction of windows in State 2. Within that state, the colorectal cancer patients displayed increased dynamic connectivity between the salience network and the right executive control network, and between the sensorimotor network and the language and primary visual networks, alongside decreased dynamic connectivity involving the left executive control network with the sensorimotor, auditory, and visual networks, and the right executive control network with the sensorimotor network. These shifting patterns suggest that the two regimens are associated not just with different average levels of connectivity but with different styles of moment-to-moment network reconfiguration.

The authors are explicit about the hierarchy of confidence in their results. The graph theory findings, having survived multiple comparison correction, are described as more robust, whereas the static and dynamic functional network connectivity results are presented as exploratory and in need of independent validation. They also stress the limits imposed by clinical reality: the study compared patients with different cancers, different metastatic patterns, and multiple regimens within each group, so it cannot isolate the pharmacological effect of carboplatin versus fluorouracil from the effects of the underlying disease. What the study does establish is that the two groups, defined by their treatment regimens, show measurably different brain network architectures at the end of therapy, a result consistent with the idea that chemotherapy-related brain changes are not uniform across drug classes.

For patients and clinicians, the significance of this work lies in its direction rather than its immediate application. The precuneus alterations associated with carboplatin-based treatment and the widespread default mode, salience, and executive network changes associated with fluorouracil-based treatment provide concrete, imaging-based targets that future longitudinal studies can follow from before chemotherapy through recovery. If larger, better-controlled cohorts confirm that specific regimens produce specific network signatures, resting-state MRI could eventually become a practical surveillance tool for the brain during cancer care, helping to identify which survivors need cognitive support and guiding the design of regimens that treat the tumor while sparing the mind. The study, approved by the Ethical Commission of Jiangsu Cancer Hospital and conducted under the Declaration of Helsinki, is open access, allowing researchers worldwide to build on its graph-theoretic and independent component analysis framework.

Subject of Research: Differential effects of fluorouracil-based versus carboplatin-based chemotherapy on resting-state functional brain networks in colorectal and lung cancer patients

Article Title: Distinct functional brain network alterations associated with 5-fluorouracil- and carboplatin-based chemotherapy regimens in CRC and NSCLC patients: a combined graph theory and group ICA study based on rs-fMRI

Article References: Distinct functional brain network alterations associated with 5-fluorouracil- and carboplatin-based chemotherapy regimens in CRC and NSCLC patients: a combined graph theory and group ICA study based on rs-fMRI. (n.d.). https://doi.org/10.1186/s12880-026-02823-0

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02823-0

Keywords: chemotherapy, cognitive impairment, resting-state fMRI, graph theory, independent component analysis, precuneus, default mode network, salience network, colorectal cancer, non-small cell lung cancer, 5-fluorouracil, carboplatin

A new study published in the journal Neuroinformatics tackles this reproducibility crisis head on. Hossein Haghighat, of the Department of Computer Engineering at Kashmar Higher Education Institute in Iran, has built a neuroinformatics framework designed to systematically compare how different functional connectivity measures perform under exactly the conditions where machine learning is most fragile: small samples. Rather than chasing another incremental gain in diagnostic accuracy, the work asks a more fundamental question, namely which mathematical descriptions of brain communication actually carry reliable information about autism, and whether the answer changes across human development.

The technical pipeline at the heart of the framework begins with group independent component analysis, a data-driven decomposition technique that separates the four-dimensional fMRI signal into spatial networks reflecting coherent, resting-state activity. Once these group-level networks are identified, dual regression is applied to extract subject-specific time series for each network in every participant. This two-step strategy, well established in the neuroimaging literature, allows each individual’s connectivity to be expressed relative to a common set of whole-brain networks, from the default mode network to attentional and sensorimotor systems, while still preserving person-level variability.

On top of these network time series, the framework computes five distinct functional connectivity measures, deliberately chosen to span the major families of interaction statistics used in the field. Full correlation captures straightforward linear co-fluctuation between networks. Partial correlation isolates direct linear relationships by statistically removing the influence of all other networks. Bivariate Granger causality introduces directionality, testing whether activity in one network helps predict future activity in another. Coherence moves the analysis into the frequency domain, quantifying synchronized oscillations at specific temporal rhythms. Finally, mutual information, an information-theoretic quantity, captures nonlinear statistical dependencies that linear measures can miss entirely. Together, these metrics cover time-domain and frequency-domain interactions, linear and nonlinear coupling, and directed and undirected relationships.

The study analyzed resting-state data drawn from the Autism Brain Imaging Data Exchange, or ABIDE, an openly shared multinational repository that aggregates scans from many imaging sites. Crucially, the analyses were stratified across three developmental stages: children, adolescents, and adults. This age-stratified design reflects a growing recognition in autism research that the brain differences associated with the condition are not static. Large-scale neural networks continue to mature throughout childhood and adolescence, and previous work by the same author and others has documented age-related patterns of both hypo-connectivity and hyper-connectivity in autism. A connectivity measure that performs well in one age band may fail entirely in another, and pooling ages can mask these developmental dynamics.

The methodological centerpiece of the framework, however, is its insistence on leakage-aware evaluation. In small-sample neuroimaging, datasets contain far more connectivity features, potentially thousands of pairwise relationships, than participants, creating a high-dimensional feature space in which classifiers can trivially overfit. The danger is compounded by a subtle but pervasive error known as data leakage, in which feature selection is performed on the entire dataset before cross-validation begins. When that happens, information from the test samples has already influenced the choice of features, inflating apparent accuracy in a way that is invisible to the researcher but catastrophic for real-world generalization. Reviews of prediction practices in psychiatry and neuroimaging have repeatedly flagged this trap as a leading cause of over-optimistic results.

Haghighat’s framework closes this loophole by performing feature selection strictly within the training folds of a leave-one-out cross-validation scheme. In every iteration of the cross-validation loop, one participant is held out, features are ranked and selected using only the remaining participants, a classifier is trained on that reduced feature set, and only then is the held-out participant classified. Multiple machine learning classifiers were employed as standardized evaluation tools, allowing the comparison to focus on the relative merits of the connectivity measures themselves rather than the quirks of any single algorithm. This disciplined protocol produces performance estimates that, while perhaps less spectacular than leaked estimates, are far more honest reflections of the information genuinely contained in each connectivity metric.

The results reveal a striking developmental structure. Linear connectivity measures, particularly full and partial correlation, showed the most stable behavior in childhood, suggesting that in young brains the dominant autism-related signal is carried by straightforward linear co-activation patterns among large-scale networks. In adolescence, by contrast, nonlinear information-theoretic measures, chiefly mutual information, proved the most informative, hinting that the reorganization of neural circuits during teenage years may generate interaction patterns that linear statistics fail to capture. In adulthood, frequency-domain measures demonstrated stronger performance, consistent with the idea that rhythmic synchronization properties of adult networks encode diagnostic information that time-domain correlation obscures. No single measure dominated across the lifespan, which is precisely the point: the optimal choice of connectivity metric depends on the developmental stage of the sample being studied.

These findings carry practical consequences for anyone building diagnostic or biomarker tools from resting-state fMRI. The autism neuroimaging community has long wrestled with the heterogeneity of the condition itself, the variability introduced by multi-site data collection, and the statistical fragility of small clinical samples. Previous multisite classification efforts have shown that reported accuracies depend heavily on sample composition, and comprehensive reviews of connectivity findings in autism have described a confusing mix of over- and under-connectivity results that defy simple summary. By benchmarking measures within a single, leakage-controlled framework and across age bands, the new study offers researchers a practical reference for selecting connectivity metrics appropriate to their populations, and a template for the kind of rigorous cross-validation that reviewers and journals are increasingly demanding.

Perhaps most importantly, the work reframes what a successful neuroimaging machine learning study should look like. Instead of presenting yet another classifier with an eye-catching accuracy figure, it emphasizes comparative methodological evaluation, transparency about overfitting risks, and developmental specificity. As the field moves toward clinical translation, where connectivity-based measures might one day support diagnosis or subtype identification, such methodological hygiene is not optional. Frameworks like this one provide the benchmarking infrastructure needed to separate genuine neural signatures of autism from statistical artifacts, and they suggest that the path to reliable neuroimaging biomarkers runs through careful, age-aware, leakage-free evaluation rather than through bigger accuracy numbers alone. The study received no external funding, and its underlying data remain publicly available through the ABIDE initiative, lowering the barrier for other teams to adopt and extend the approach.

Subject of Research: Evaluation of functional connectivity metrics for machine learning analysis of resting-state fMRI in age-stratified autism spectrum disorder research

Article Title: A Neuroinformatics Framework for Evaluating Functional Connectivity Metrics in Small-Sample Resting-State fMRI: An Age-Stratified Autism Study

Article References: Haghighat, H. (2026). A Neuroinformatics Framework for Evaluating Functional Connectivity Metrics in Small-Sample Resting-State fMRI: An Age-Stratified Autism Study. Neuroinformatics, 24(3), Article 60. https://doi.org/10.1007/s12021-026-09816-y

Image Credits: AI Generated

DOI: 10.1007/s12021-026-09816-y

Keywords: autism spectrum disorder, functional connectivity, resting-state fMRI, machine learning, independent component analysis, dual regression, Granger causality, mutual information, coherence, cross-validation, data leakage, neuroinformatics