Biology subjects

Shen, L.

Publications and source records attributed to Shen, L..

At least 19 recordsLinked to original sources

bioRxiv2026

EG-BINN: Parameter-Efficient Ensembling of Gated Biologically Informed Neural Networks for Epigenetic Age Estimation

Epigenetic aging clocks are widely used to evaluate biological aging, disease risk, and clinical interventions. However, most existing clocks rely on linear models that overlook nonlinear relationships between age and methylation patterns, or on deep learning architectures that are parameter-intensive and difficult to interpret. We developed EG-BINN, a parameter-efficient ensemble of biologically informed neural networks that uses sample-specific gating to modulate representations across biologically annotated nodes. Trained on 93 multi-tissue datasets, EG-BINN achieved the lowest mean and median absolute errors among matched linear and nonlinear baselines in five-fold cross-validation, as well as the lowest mean absolute and root mean squared errors across eight held-out Gene Expression Omnibus (GEO) datasets. Analysis of the gating mechanism revealed widespread associations between pathway gate values and age, including trajectories that varied across blood, adipose tissue, skeletal muscle, and liver. Across 11 TCGA cohorts, EG-BINN identified significant tumor--normal differences in residual epigenetic age acceleration in six cohorts, whereas Shapley additive explanations (SHAP) highlighted recurrent pathway nodes contributing to tumor age predictions. EG-BINN combines accurate pan-tissue age prediction with an explainable architecture for examining biologically labeled nodes across age, tissue, and disease contexts.

bioinformatics↗

bioRxiv2026

Integrating Multisource Target Evidence with Machine-Learning Potency Prediction for Alzheimer's Disease Drug Discovery

Alzheimer's disease (AD) target evidence is distributed across biological, chemical, structural, regulatory, and trial-registry sources, complicating reproducible target evaluation and modeling. We developed a traceable data-to-model platform that integrates AGORA-centered evidence through Target Explorer, Drug-Target Lookup, and Model Portfolio workspaces. The resource contains 967 targets, 452,249 compounds, and 1,430,260 measurements. On a curated PIK3CA dataset, target-specific fine-tuning of the Boltz-2 affinity head increased Pearson r from 0.248 to 0.560 and reduced mean absolute error from 0.900 to 0.706 on 469 held-out compounds. Separately, we screened 42 regression models over 50 random splits for ten moderate-sized target datasets; a random forest model for XDH achieved the highest mean R^2 (0.815). The platform connects curated evidence, model-ready datasets, reproducible evaluation, and public results while clearly separating contextual evidence from therapeutic claims and random-split performance from generalization to unseen chemical scaffolds.

bioinformatics↗

bioRxiv2026

A public-data reproducibility benchmark of a 2.5-D U-Net with a slice-level topology term for type-B aortic dissection segmentation

Background: Type-B aortic dissection requires separation of the true and false lumina in heterogeneous, anisotropic CTA. A topology term may reduce fragmentation, but its reproducibility must be tested under a fixed case split and independent native-grid evaluation. Methods: We used the public Graz type-B aortic dissection CTA collection (40 cases; CC BY 4.0) and froze a 28/6/6 case split for training, development, and held-out testing. A compact 2.5-D U-Net trained with cross-entropy plus multiclass Dice was compared with the identical model augmented by a fixed-weight (0.5) differentiable slice-level soft clDice term. Two training seeds were run for each arm. All 40 cases passed the NRRD audit; predictions for six held-out cases were restored to their native grids and scored with case-level 3-D voxel-grid Dice, voxel-grid morphological clDice, 6-connected component counts, physical-volume error, and 2,000-resample case-level paired bootstrap intervals. Results: With seed 20261007, topology-minus-baseline differences were -0.028 (95% CI -0.053 to -0.004) for true-lumen Dice, +0.032 (-0.006 to +0.096) for true-lumen voxel-grid clDice, +0.071 (+0.037 to +0.103) for false-lumen Dice, and +0.073 (+0.048 to +0.103) for false-lumen voxel-grid clDice. With seed 20261008, the corresponding differences were -0.040 (-0.065 to -0.018), +0.013 (-0.034 to +0.071), -0.099 (-0.187 to -0.004), and -0.058 (-0.124 to -0.001). Descriptive pooled means over 12 case rows per arm were true-lumen Dice 0.577 for topology versus 0.611 for baseline and false-lumen Dice 0.409 versus 0.422; these pooled rows are not independent cases. Conclusions: Under this fixed six-case test split, the corrected slice-level topology term did not produce a consistent benefit. True-lumen Dice was lower in both seeds, while false-lumen effects reversed direction between seeds; topology-clDice changes were small and/or interval-crossing. The benchmark supports an auditable failure-analysis workflow but does not support a general topology benefit, clinical use, external generalization, or a panvascular decision system.

bioinformatics↗

bioRxiv2026

Panvascular-JEPA: Unseen-Structure Evaluation of Factorized Predictive States for Hemodynamic Networks

Predictive representations for vascular simulation should be tested on unseen network structures and compared with the governing equations before being interpreted as general-purpose world models. We constructed a new, dimensioned but uncalibrated nodal resistance-inertance-compliance benchmark containing all 38 connected, non-isomorphic six-node graphs with five to seven edges. Twenty-four structures supplied training, validation, and new-background testing; 14 structures were wholly withheld. Across 648 parameter backgrounds and two prescribed action programs, 1,296 trajectories contained 96 forecast steps over three cycles. Four graph predictors received equal six candidate optimization searches, followed by five fixed-seed fits. Selected configurations and analysis code were locked before test trajectories were generated. Comparators included known-equation numerical integrators, an inertia-free reduction, and a five-coefficient model fitted to training trajectories. Unseen-structure mean standardized RMSE was 8.46 x 109 for four-factor OPF, 329 for single factor OPF, and 1.83 x 1031 for supervised graph prediction, compared with 1.09 x 10-4 for foursubstep known-equation integration and 6.77 x 104 for the fitted mechanism. The primary OPF-4 minus OPF-1 difference was 8.46 x 109, with a paired stratified-graph 95% percentile interval of [2.23 x 105, 2.54 x 1010]. Neural errors were strongly dependent on training seed and network structure, with extreme forecasts dominating the means. All four neural means exceeded persistence; relative performance within this comparison did not establish a usable forecasting model. The study supplies a distinct numerical test of structure transfer, factor-count control, and recursive stability. Its conclusions concern the implemented models, optimization budget, and finite graph catalogue. It establishes neither patient-specific performance nor anatomical transfer, biological remodeling, device efficacy, or clinical readiness.

bioengineering↗

bioRxiv2026

Towards Robust Automated Mouse Scratching Behavior Detection for Itch Research

Automated detection of mouse scratching behavior could reduce manual annotation and is essential for preclinical itch research. However, robust performance across laboratories remains a challenge for current methods. We present Scratch-AID+, a framework dedicated to automated mouse scratching detection, including a diverse and progressive set of approaches: classic ResNet-based detectors, task-adaptive fine-tuning of video foundation models, and in-context learning with advanced multimodal large language models. To simulate real-world use cases, we benchmark on a public dataset and utilize additional cross-laboratory video recordings for independent testing. We find that classic baselines suffered severe performance degradation on external data under the realistic setting with sparse scratching events. In contrast, fine-tuned video foundation models successfully bridged this generalizability gap, achieving significantly better detection performance. Furthermore, frozen multimodal large language models demonstrated promising out-of-the-box ability using only few-shot demonstrations. We also examined the effects of different video preprocessing strategies. Our findings highlight the improved performance and generalizability of the adapted foundation models for automated scratching behavior analysis and shed light on the real-world deployment of such detectors. Our code is publicly available at https://anonymous.4open.science/r/scratch aid-plus/.

bioinformatics↗

bioRxiv2026

PanVasc Research for AI assisted evidence analysis in panvascular intervention

Panvascular intervention research requires evidence workflows that preserve source identity, outcome definitions and observation windows. We developed PanVasc Research, an executable research framework, and evaluated a fixed local Qwen3-4B model using two complementary tasks. Fifty ClinicalTrials.gov records from five vascular query strata generated 200 source-fidelity tests with intact evidence or controlled removal of the requested source, primary outcome or timeframe, plus 50 clean controls. A separate 100-statement sample from the official NLI4CT test set assessed clinical-trial entailment and evidence selection against existing expert labels. Generic and checklist prompts used identical evidence and maximum generation budgets; each output was also evaluated with an input-only deterministic contract. Registry exact accuracy was 51/200 (25.5%) with the generic prompt and 77/200 (38.5%) with the checklist; intact-case agreement was 50/50 and 48/50. The rule baseline recovered 200/200 tuples. NLI label accuracy was 48/100 (48.0%) and 50/100 (50.0%), respectively. The source contract retained 35 and 31 incorrect NLI labels in the two arms. Registry references establish fidelity to a registration snapshot, while NLI4CT concerns breast-cancer trials and does not validate vascular expertise. The framework also retains provenance-recorded literature retrieval, structured research planning and local numerical analysis. These experiments support a bounded assessment of source handling and semantic failure, rather than a new foundation model or autonomous scientific discovery. A proposed endpoint representation identifies the additional domain annotation and independent validation required for a panvascular research model.

bioinformatics↗

bioRxiv2026

KMT2E recruitment by H3 serotonylation regulates neurodevelopmental chromatin dynamics

Histone H3 Gln 5 serotonylation (H3Q5ser) is a recently identified epigenetic modification in brain that modulates reader interactions with adjacent H3 Lys 4 trimethylation (H3K4me3) to promote transcriptional permissiveness. However, whether H3K4me3Q5ser and its associated binding proteins regulate neurodevelopmental gene expression programs remains unknown. Here, we identified the catalytically inactive Lysine methyltransferase 2E (KMT2E) as a reader of combinatorial H3K4me3Q5ser. KMT2E preferentially binds H3K4me3Q5ser over H3K4me3 alone, and enriches at broad chromatin domains marking actively transcribed neurodevelopmental loci. Notably, heterozygous variants in KMT2E have been implicated in ODLURO syndrome, a recently characterized neurodevelopmental disorder (NDD). To identify the molecular mechanisms underlying ODLURO syndrome, we generated a Kmt2e transgenic mouse model that reproduces behavioral, physiological, and cellular endophenotypes associated with this and other NDDs. Furthermore, we observed that KMT2E mediates these effects by recruiting the NCoR/HDAC3 repressor complex to H3K4me3Q5ser-marked loci to restrict spreading of co-localized H3 Lys 9 acetylation (H3K9ac). Inhibition of aberrant H3K9ac spreading was sufficient to rescue transcriptional dysregulation in Kmt2e haploinsufficient neurons. These findings thus establish KMT2E as a critical reader of H3 serotonylation during neurodevelopment and provide mechanistic insights into the pathogenesis of ODLURO syndrome.

neuroscience↗

bioRxiv2026

TEDDY: An integrative workflow for TE-chimeric isoform reconstruction and systematic characterization of TE-dependent transcriptional regulation

Transposable elements (TEs) substantially expand transcriptional regulation and complexity. Despite advances in sequencing technologies, the systematic interrogation of TE-dependent transcripts across diverse samples and platforms remains limited. Here we introduce TEDDY, a computational workflow for large-scale reconstruction, quantification and regulatory inference of TE-dependent isoforms. Using TEDDY, we resolved the landscape of TE-dependent isoforms in mammalian preimplantation development, revealing a conserved pattern of stage-specific isoform generation via species-specific TE exonization. We reconstruct a TE-dependent regulatory network underlying pluripotent-to-totipotent transition wherein key transcription factors are driven by TE-derived promoters and validated experimentally. Within this network, Arid3a, a novel regulator nominated by TEDDY, was functionally validated as essential for totipotency establishment and early development. Further application of TEDDY to hepatocellular carcinoma identified recurrence-associated prognostic isoforms, underscoring their clinical relevance. Benchmarking establishes TEDDY's accuracy and efficiency with unique capabilities in full-length isoform recovery, cross-sample and -platform analysis, visualization, and TF-TE-gene network reconstruction, making TEDDY applicable across diverse biological contexts.

genomics↗

bioRxiv2026

Deconvolution-derived cell-type expression targets for personal genome sequence-to-expression prediction

Sequence-to-function models learn regulatory features from genomic sequence, but they remain limited in their ability to predict gene-expression differences among individuals. Cell-type-specific regulatory effects may be obscured in bulk RNA sequencing, whereas paired genotype and single-cell expression cohorts remain small. We evaluated whether deconvolution of bulk RNA-seq could provide scalable cell-type-specific targets for personal-genome expression prediction. GTEx v8 bulk RNA-seq from six tissues was deconvolved with BayesPrism using single-nucleus reference profiles, producing targets across 83 tissue-cell-type contexts. Deconvolved expression agreed with matched pseudobulked GTEx single-nucleus RNA-seq, with median donor-level Pearson correlations across genes ranging from 0.53 to 0.73 by tissue. We compared genotype-feature models, regressors trained on frozen Enformer representations, and fine-tuned Enformer and Borzoi models. Across random and nonlinear-enriched gene sets, sequence-derived approaches generally outperformed genotype-feature baselines, while frozen Enformer features were competitive with end-to-end fine-tuning. For the random gene set, Fisher-averaged Pearson correlations were 0.122-0.142 for sequence-derived approaches and 0.081-0.086 for genotype-feature baselines in a coverage-aware sensitivity analysis. Model performance was positively associated with deconvolution-pseudobulk agreement for sequence-derived models (r = 0.35-0.43 across tissue-cell-type contexts), suggesting that target reliability may constrain downstream prediction. Context-specific Enformer fine-tuning did not materially out-perform a shared, combined-context strategy. These results support deconvolution as a feasible approach for generating cell-type-resolved training targets, while showing that target quality and limited cohort size remain important constraints. Frozen pretrained representations provide a computationally efficient and competitive baseline for personal sequence-to-expression modeling.

bioinformatics↗

bioRxiv2026

Single-cell foundation models predict durable CAR T response despite imperfect cell annotation

CD19-targeted chimeric antigen receptor (CAR) T cell therapy achieves high initial response rates in B-cell acute lymphoblastic leukemia (B-ALL), yet half of patients relapse within one year. Pre-infusion product composition decoded by single-cell RNA sequencing (scRNA-seq) carries information predictive of long-term CAR T persistence, but extracting this information from individual patients typically requires highly sophisticated bioinformatics expert annotation, limiting clinical translation. Here, we evaluate whether single-cell foundation models (scFMs) can extract clinically actionable information from engineered CAR T products. We applied four scFMs (scGPT, scFoundation, CellPLM and UCE), including fine-tuned versions of scGPT and scFoundation, to paired basal and CD19-stimulated pre-infusion CAR T products from 33 pediatric patients with B-ALL. Although annotation accuracy declined relative to healthy peripheral blood references, scFM-derived cell composition stratified patients with long-duration B-cell aplasia with a leave-one-out cross-validated area under the receiver operating characteristic curve of 0.879 (95% confidence interval, 0.742-0.986). Notably, foundation-model-identified cell proportion analysis matched or exceeded expert annotations for several predictive features, demonstrating that accurate clinical prediction may not require perfect per-cell annotation to begin with. CD8+XCL1/2+ cells were further identified as the biomarker consistently associated with durable CAR T persistence across models under CD19 stimulation, whereas other candidate populations showed limited reproducibility. Finally, we translate these findings into a locally deployable decision-support AI agent that predicts the probability of sustained CAR T persistence from pre-infusion CAR T scRNA-seq data.

systems biology↗

bioRxiv2026

Integrative single-nucleus multi-omics profiling identifies candidate regulators and signaling axes in Alzheimer's disease lipid-processing microglia

Lipid-processing microglia are among the microglial states most strongly associated with Alzheimers disease (AD) pathology, yet whether this association reproduces across independent cohorts, what transcriptional programs define the state, and which upstream signals and small molecules can modulate it remain unsolved. We address these questions through a cross-cohort analysis of one such substate (MG4) by integrating differential expression, transcription factor activity inference, gene set enrichment, and cell-cell communication across five independent single-nucleus RNA sequencing cohorts (ntotal = 140 donors), with paired single-nucleus ATAC sequencing in one multi-omic cohort for epigenomic corroboration. A held-out cohort (n = 150 donors) supported donor-level regression of MG4 proportion on ligand expression, and two spatial transcriptomics datasets (ntotal = 30 donors) related ligand expression to MG4 identity in neighboring spots. MG4 was reproducibly enriched in AD across all five cohorts (pooled log2 fold change = 0.90, p = 3.0 x 10-4). Expression-based inference and motif accessibility jointly nominated MITF and BACH1 as regulators of a program led by V-ATPase-driven lysosomal acidification and cholesterol efflux, a lysosomal-biogenesis signature distinct from the catabolic DAM and lipid-storage LDAM programs, with AD-specific upregulation of energy metabolism. FGF1 and TGFB2 were the most supported candidate upstream ligands, each significant in donor-level regression with further spatial evidence. Computational drug repurposing nominated ten blood-brain barrier-penetrant compounds as perturbational probes. Together, these results advance a described disease-associated microglial state into a reproducible, mechanistically framed regulatory model, providing candidate regulators, upstream ligands, and pharmacological probes for functional validation.

bioinformatics↗

bioRxiv2026

Learning Shapes the Energy Cost of Neural Tasks

The stark difference in energy use between AI systems and biological brains highlights both the remarkable efficiency of the brain and our limited understanding of the energetics underlying neural computation. Although efficiency is widely cited as a principle of neural design, direct measurement of the energy cost of specific neural tasks within their corresponding circuits has remained limited. Here, by simultaneously measuring intracellular glucose and calcium dynamics in the same behaving mice in vivo, we use neuronal glucose consumption to define circuit-level energy costs associated with learning-based behavioral tasks. We found that the post-learning fuel cost per task was significantly lower than pre-learning levels across multiple hippocampus- and cortex-dependent learning models, with or without changes in bulk calcium dynamics. This change in fuel cost was not mediated by extracellular glucose transport but instead reflected a reduction in intracellular glucose consumption and depended on canonical plasticity mechanisms, including NMDAR signaling and protein synthesis. Together, these findings suggest that attenuation of task-specific energy cost may represent a general bioenergetic trajectory of learning and plasticity. This "energy minimization" hypothesis provides insight into how the biological brain achieves efficiency and offers an orthogonal, complementary perspective to neural activity-centered frameworks.

neuroscience↗

bioRxiv2026

Intravital imaging of age-related conjunctival lymphatic changes on the ocular surface

PurposeLymphatic system plays a critical role in fluid regulation and immune response. The conjunctiva of the ocular surface is endowed with a rich lymphatic network, and it provides an ideal site to study lymphatic structure and function. The purpose of this study is to investigate potential morphological and functional changes of conjunctival lymphatics with aging, a time-dependent biological deterioration process. MethodsYoung and aged fluorescently labeled transgenic mice of Prox-1, the master control gene for lymphatic determination, were used in the study. For morphological assessment, conjunctival lymphatic vessels were examined in vivo by our advanced live imaging system. For functional analysis, lymphatic drainage efficiencies were measured by fluorescently labeled tracer injection. ResultsCompared to young mice, both vascular branching points and intraluminal valves were significantly reduced in conjunctival lymphatic vessels of aged mice. Moreover, lymphatic functional deterioration and drainage deficiencies, such as fluid leakage and reflux, were also detected in the aged condition. ConclusionsConjunctival lymphatic system undergoes morphological as well as functional changes with aging. Further investigation into this phenomenon may provide novel insights into lymphatic and age-related diseases inside and outside the eye.

immunology↗

bioRxiv2026

Killer-cell dominance dichotomy governs tumor immune networks and stratifies inflamed cancers

Cancer immunotherapy benefits remain limited, even among "hot" tumors with high killer lymphocyte infiltration. Here, we investigated the population-level architecture of killer cells based on nearly 5,000 pan-cancer scRNA-seq samples, together with orthogonal validation by spectral cytometry and spatial transcriptomics. Unlike the prevailing "hot-cold" paradigm, which assumes coordinated infiltration of multiple cytotoxic lineages, we uncovered a conserved framework wherein terminal cytotoxic immunity in individuals or malignancies diverges into states dominated by either exhausted CD8+ T cells (Tex) or CD56dimCD16hi NK (NK1) cells. Despite the complexity of the tumor microenvironment, Tex-NK1 divergence governs the primary axis of tumor-intrinsic and tumor-extrinsic variance. Distinct from the conventional view that NK cells positively contribute to immunotherapy efficacy, NK1-skewed tumors, although highly cytolytic, are refractory to current immune checkpoint blockade regimens. This killer divergence defines a foundational axis of cancer immunity and provides a resource for prioritizing next-generation targets for NK-directed immunotherapy.

immunology↗

bioRxiv2026

BCMA/CD19 dual-targeting with FasTCAR: AZD0120 demonstrates potent preclinical efficacy in multiple myeloma

Chimeric antigen receptor (CAR)-T cell therapy has improved outcomes for patients with multiple myeloma (MM), but its broader use is restricted by manufacturing complexities and treatment-related toxicities. AZD0120 is a dual-targeting B-cell maturation antigen (BCMA)/CD19 CAR-T cell therapy manufactured via the rapid FasTCAR process. We developed a dual-targeting "loop" CAR that incorporates a novel humanized anti-BCMA single-chain variable fragment (scFv), clone SG, and an FMC63-derived anti-CD19 scFv. This AZD0120 CAR preserved functional binding to both antigens and conferred robust in vitro and in vivo cytotoxicity while maintaining single-antigen reactivity. Conventional manufacture of CAR-T cells with the AZD0120 CAR (AZD0120C) yielded cells with minimal tonic signaling, limited responsiveness to soluble BCMA, and preservation of naive/stem cell memory-enriched phenotypes, yet robust cytokine production upon BCMA+ target engagement. AZD0120C demonstrated cytotoxicity comparable to benchmark BCMA CAR-Ts across MM lines in vitro and showed strong in vivo expansion and tumor control in xenograft models. FasTCAR manufacturing - designed to shorten vein-to-vein timelines and enrich less-differentiated phenotypes - further enhanced in vivo performance: AZD0120 consistently achieved superior tumor control and greater CAR-T expansion vs AZD0120C across disseminated MM.1S, NALM-6, and JeKo-1 models, with superior efficacy observed at lower cell doses. Collectively, these data support clinical evaluation of AZD0120 as a differentiated BCMA/CD19 CAR-T cell therapy with the potential to improve disease control and patient access in MM. Key PointsO_LIAZD0120 is a dual-targeting CAR-T that displays a favorable anti-myeloma functional profile and co-targets a source of potential relapse C_LIO_LIThe FasTCAR process yields TN/SCM-rich CAR-T populations, promotes in vivo expansion and achieves potent tumor control in xenograft models C_LI

cancer biology↗

bioRxiv2026

Repurposing of the assembly chaperone PAC1-PAC2 defines a specialized proteasome landscape in mature mammalian sperm

Mature mammalian sperm rely on active proteasomes containing the testis-specific subunit 4s for capacitation, acrosomal exocytosis and fertilization, but how these proteasomes are structurally specialized remains unclear. We determine the endogenous proteasome landscape of mature bovine spermatozoa and testis using cryo-electron microscopy. We resolve five sperm proteasome assemblies: PA200-20S, free 20S, a previously unrecognized PA200-20S-PAC1/2 hybrid, and PAC1/2-20S complexes capped at one or both ends. Unexpectedly, the PAC1/2-bound complexes are mature and POMP-free, selectively incorporated for 4s, lack 2, and adopt a PAC1/2 orientation and gate configuration distinct from known assembly intermediates. Together with elevated proteolytic activity, these features indicate that PAC1/2 is repurposed from a transient assembly chaperone into a stable post-assembly regulator in mature sperm. 4s-containing proteasomes assemble through the canonical PAC1-PAC4/POMP pathway yet may follow {beta}2-{beta}3-{beta}1 incorporation ordering; functionally, 4s incorporation remodels the sCP surface and enhances peptidase activities, definining the spermatoproteasomes as a hyperactive 20S variant. PAC1/2-bound proteasomes are absent from testis, where PA200-20S and 26S proteasomes are present, while 26S proteasomes are nearly lost in mature sperm. Our findings define a sperm-specific, ATP-independent proteasome landscape, expand the functional repertoires of 4s and PAC1/2, and provide a structural framework for proteasome specialization in male fertility.

cell biology↗

bioRxiv2026

Beyond Phase Estimation: A Multidimensional Gating Framework for Robust Real-Time Closed-Loop Neural Stimulation

Neural oscillatory phase is widely used as a control variable in real-time closed-loop stimulation, yet its validity under strict causal constraints and noisy conditions has rarely been systematically examined. We introduce a Multidimensional Gating Framework (MGF), a plug-in and estimator-agnostic module that determines whether phase information should be admitted into control by evaluating instantaneous amplitude, narrowband signal-to-noise ratio (SNR), and spectral peak ratio (PR) within a strictly causal window. Using causal streaming replay on a public resting-state EEG dataset, we benchmarked Hilbert based phase estimation and endpoint-corrected Hilbert estimation with and without MGF. Among feasible subjects, MGF significantly reduced phase dispersion for both estimators, while robustly suppressing catastrophic phase errors. In contrast, ungated approaches exhibited systematic failures under the same conditions.

neuroscience↗

bioRxiv2026

Faithful Supervised Dimensionality Reduction for Biomedical Data via Decision Geometry

Unsupervised dimensionality reduction methods aim to preserve intrinsic data geometry by maintaining local neighborhoods and approximate global relationships in low-dimensional embeddings, but they do not use label information and therefore may fail to reflect task-relevant class structure in biomedical and health applications. Supervised dimensionality reduction (SDR) incorporates labels to improve class organization, yet existing approaches often face a trade-off between discrimination and geometric faithfulness. Linear supervised methods are stable and interpretable but are limited in their ability to capture nonlinear structure, whereas many nonlinear methods impose supervision directly in the embedding space, which can over-separate classes and distort the underlying manifold. In biomedical applications, labels such as cell types in single-cell data or patient status in clinical cohorts provide meaningful biological signal, and supervised dimensionality reduction can use this information to produce more informative low-dimensional representations. Here we propose a new framework, DG-UMAP (Decision-Geometry UMAP), for faithful supervised dimensionality reduction via decision geometry. We first fit a classifier in the original feature space and use its boundary-local decision geometry to construct a low-rank metric deformation that emphasizes discriminative directions while limiting geometric distortion. Parametric UMAP is then applied to the transformed space, so supervision acts through the ambient geometry rather than by directly forcing class separation in the embedding. Across synthetic and multiple real-world biomedical datasets, our method yields embeddings with improved agreement with class structure and global organization while preserving local neighborhood quality.

bioinformatics↗