A global home for research

Lipid droplets (LDs) are dynamic organelles that serve as reservoirs for neutral lipids. Lipids are indispensable building blocks of all cells, forming the membranes that protect cells and the organelles within. The human brain is a lipid-rich organ; the distribution and trafficking of lipids inside brain cells are both functionally vital for neuronal insulation and signaling and pathologically significant. Because lipid misregulation has been linked to neurodegenerative diseases such as Alzheimer’s disease (AD), accurate visualization of LDs is critical for uncovering lipid trafficking defects that may contribute to AD onset. This study compared two fluorescent LD-labeling dyes, BODIPY 493/503 and LipidSpot 610, to evaluate whether the newer dye, LipidSpot 610, provides equal or improved specificity for LD labeling compared to the standard of BODIPY 493/503. Using optimized loading conditions, 3T3 fibroblast cells were stained and imaged via confocal microscopy. Lysosomal markers were used to ensure that each dye specifically labeled LDs rather than lysosomes. Image analysis revealed that both dyes specifically and accurately stained LDs. Interestingly, the lower Pearson’s correlation coefficient observed in the LipidSpot 610 condition suggests that it may provide greater LD-labeling specificity than the standard LD dye, BODIPY 493/503. Furthermore, the colocalization of BODIPY 493/503 and LipidSpot 610 confirms that both dyes accurately target the same intracellular LD structures. By evaluating the specificity and accuracy of both LD-labeling dyes, this study supports future research aimed at understanding lipid trafficking and its role in cellular metabolism and neurodegenerative diseases like AD.

NSRI-J-2026-0215 · Full text forthcoming

Background. Fish swimming in natural environments often exploit vortices shed by nearby objects or other fish to improve their propulsive efficiency. A notable example of this phenomenon is the Kármán gait, where fish synchronize their body undulations with vortex streets behind stationary obstacles to attain hydrodynamic benefits. This study investigated whether a bioinspired robotic fish, the Knifebot, could replicate these hydrodynamic mechanisms. Methods. Experiments were conducted using both a scaled resin model and a full-scale robotic Knifebot in flow tanks with controlled vortex shedding. The resin model was tested across multiple flow conditions to characterize passive vortex wake interactions, while the full-scale Knifebot was evaluated using high-speed video, LED tracking, onboard inertial sensors, and particle image velocimetry (PIV). Motion data were processed using MATLAB to quantify robot kinematics, acceleration, vortex-phase synchronization, and wake structure. Results. The results identified a wake-capture region in which the resin model exhibited pronounced wake-coupled motion and upstream streamwise displacement. During a 7-s trial, the model advanced approximately 0.364 m toward the stationary obstacle against the imposed flow, reached velocities up to 0.665 m/s, and exhibited acceleration up to 0.304 m/s2. The PIV analysis confirmed strong vortical structures within the same wake-capture region. Full-scale experiments further demonstrated sustained wake-coupled behavior, including elevated fin response and repeated reversals in accumulated yaw error that limited persistent directional drift. Together, these results show that obstacle-generated vortex wakes can produce kinematic and directional responses under favorable wake-coupling conditions. Conclusion. By bridging fluid dynamics and robotics, this research provides experimental evidence that bioinspired robots can exhibit wake-capture behaviors consistent with vortex exploitation observed in biological swimmers. Such insights may inform the design of autonomous underwater vehicles with improved hydrodynamic efficiency.

NSRI-J-2026-0211 · Full text forthcoming

Background: Neural ordinary differential equation (Neural ODE) classifiers depend on numerical solvers during inference. Strict adaptive solvers can provide highly accurate numerical solutions but may require substantially more computation than fixed-step solvers. Low-cost solvers are attractive for deployment, but their predictions can occasionally differ from a high-accuracy numerical reference. Methods: We developed CSDC-NODE, a selective inference framework that compares coarse and step-doubled Runge–Kutta solutions to measure numerical discrepancy. A residualized split-conformal procedure, calibrated on independent data, converts this discrepancy into an upper bound on reference-logit error. The coarse prediction is accepted only when its logit margin exceeds twice the calibrated bound; otherwise, the method falls back to a strict adaptive reference solver. The frozen primary policy was evaluated across 16 pre-declared runs on Moons, Spiral, MNIST, and FashionMNIST. Results: The policy preserved reference-level accuracy on Moons, MNIST, and FashionMNIST and differed by only 0.0002 on Spiral. Mean equivalent-NFE savings were 84.96% for Moons, 60.29% for Spiral, 80.02% for MNIST, and 72.68% for FashionMNIST. Step-doubling discrepancy showed strong association with reference-logit error, with mean Spearman correlations from 0.9728 to 0.9997. Empirical coverage remained close to the nominal 0.95 target, while false-certification frequencies were zero or near zero in the reported evaluations. Conclusion: CSDC-NODE provides a practical way to reduce Neural ODE inference cost while selectively preserving agreement with a strict numerical reference.

NSRI-J-2026-0209 · Full text forthcoming

Volatility forecasts are central to quantitative risk management, portfolio construction, derivative valuation, and volatility-targeting strategies. This study evaluates the one-step-ahead forecasting performance of exponentially weighted moving average (EWMA) and generalized autoregressive conditional heteroskedasticity GARCH(1,1) models in a controlled Monte Carlo framework. Three data-generating environments representing lower, moderate, and high volatility persistence are simulated using Student-t innovations. Each simulated return path contains 1,800 observations, with the first 1,000 used for estimation and the remaining 800 reserved for out-of-sample evaluation. Forecast accuracy is assessed using mean squared error (MSE), mean absolute error (MAE), and quasi-likelihood loss (QLIKE). The results indicate that model performance depends on the persistence structure and the alignment between the fixed EWMA decay parameter and the underlying volatility process. Estimated GARCH(1,1) can outperform a fixed-lambda EWMA specification when the underlying persistence differs materially from the imposed EWMA decay assumption, while EWMA remains competitive when its decay parameter is well matched to the volatility dynamics. The study is explicitly methodological and simulation-based; it does not claim empirical evidence from U.S. equity markets. The contribution is a transparent and reproducible framework for comparing volatility forecasts under known conditional-variance dynamics and for separating simulated evidence from future empirical validation.

NSRI-J-2026-0199 · Full text forthcoming