The predictive performance of molecular machine learning models depends on how molecules are featurized, yet fair comparisons across methods are rare because studies often vary architectures and hyperparameters. We benchmark five geometry-based molecular featurization techniques that operate on three-dimensional atomic coordinates, i.e., Coulomb Matrix, Bag of Bonds, Atom-Centered Symmetry Functions (ACSF), Smooth Overlap of Atomic Positions (SOAP), and Weighted Views, across five benchmark datasets (ESOL, FreeSolv, BBBP, BACE, ClinTox). While Weighted Views employs a unique weighted aggregation architecture, we controlled for model capacity by adjusting network widths so that all methods had roughly the same number of trainable parameters. All representations are evaluated under stratified k-fold cross-validation with three independent seeds, and results are reported as the mean and standard deviation across the resulting test folds.
The sound horizon scale rs is a key source of information for measurements of H0 from early-time data, and is therefore a common target of new physics proposed to solve the Hubble tension. We present a sub-2% measurement of the Hubble constant that is independent of this scale, using data from the first data release of the Dark Energy Spectroscopic Instrument (DESI DR1). Building on previous work, we remove dependency on the sound horizon size using a heuristic rescaling procedure at the power spectrum level. A key innovation is the inclusion of uncalibrated (agnostic to rs) post-reconstruction BAO measurements from DESI DR1, as well as using the CMB acoustic scale theta & lowast; as a high-redshift anchor. Uncalibrated type-Ia supernovae are often included as an independent source of ohm m information; here we demonstrate the robustness of our results by additionally considering two supernova-independent alternative datasets. We find somewhat higher values of H0 relative to our previous work: 69.2+1.3-1.4, 70.3+1.4-1.2, and 69.6+1.3-1.8 km s-1 Mpc-1 respectively when including measurements from i) Planck/ACT CMB lensing & times; unWISE galaxies, ii) the DES Year 3 6 & times;2pt analysis, and iii) Planck/ACT CMB lensing + the DES Year 5 supernova analysis. These remarkably consistent constraints achieve better than 2% precision; they are among the most stringent sound horizon-independent measurements from LSS to date, and provide a powerful avenue for probing the origin of the Hubble tension.
The late Campanian nodosaurid Panoplosaurus mirus Lambe 1919 was named on the basis of a complete skull and partial postcranium. However, as is typical of ankylosaurs, the constituent elements of the holotype skull (CMN 2759) are obscured both by their extensive co-ossification and by the overlying bony caputegulae, complicating taxonomic referral of other material. The holotype skull is re-examined here using computed tomography (CT) imaging, revealing new information about the internal anatomy. Previously unknown regions of the skull are described, including the basicranium and palate, and the internal anatomy in Panoplosaurus is examined from CT-rendered models of the brain and nasal cavities. The occurrence of a vomerine canal or groove and convoluted nasal passages are confirmed in Panoplosaurus, while decreased brain flexure relative to other nodosaurids suggests there is greater variability in the paleoneurology of Nodosauridae than previously appreciated. The unique pattern of cranial ornamentation in Panoplosaurus is also determined to be taxonomically significant (as it is in Ankylosauridae), and the generic separation of Panoplosaurus and the closely related Edmontonia is supported. Further, intraspecific variation in the cranial ornamentation of Panoplosaurus is revealed, based on the presence or absence of the internasal caputegulum, although the nature of this variation (e.g., ontogeny, sexual dimorphism, individual variation) remains elusive.
Efficient, effective, and equitable disaster response is critical yet challenging due to uncertainties, particularly those in vehicle travel times caused by infrastructure damage. To address this, current study proposes a scenario-based robust optimisation social cost vehicle routing problem (SRO-SCVRP) that minimises three objectives: (1) social cost, balancing logistics and deprivation costs, with the latter quantifying survivors' suffering from delayed access to critical supplies; (2) the variation in expected deprivation cost across travel time scenarios, reflecting solution robustness; and (3) penalties for exceeding service coverage windows (SCW), which are thresholds set by decision-makers to ensure timely deliveries, thereby enhancing model robustness. Logistics and deprivation costs operationalise efficiency and effectiveness, respectively, while the second and third objectives promote equity by reducing disparities in access and ensuring the timely fulfilment of demand. Numerical results from the SRO-SCVRP highlight a trade-off wherein gains in robustness and equity are achieved with only a modest impact on logistics cost. To assess robustness and equity, a Time and Inventory Service-Level Metric (TISM) is developed and applied to alternative water distribution strategies. Among them, the hybrid strategy yields the highest overall TISM, making it the most robust and equitable option across varied SCW thresholds and resource availability.
We lack a mechanistic understanding of how cortical contributions to balance control change in aging and Parkinson's disease (PD). Balance is governed by brainstem circuits, with higher-order centers like the cortex or basal ganglia becoming engaged as challenge increases or balance health declines. We previously showed that parallel sensorimotor feedback loops engaging brainstem and cortical circuitry contribute to muscle activity for balance control in young adults (YAs). Here, we analyze data from male and female older adults (OAs) with and without PD, decomposing perturbation-evoked tibialis anterior and medial gastrocnemius muscle activity into hierarchical components based on latencies of feedback control loops. We found that balance-correcting muscle activity followed a stereotypical waveform of long-latency responses (LLRs): LLR1 began similar to 120 ms and LLR2 occurred similar to 210 ms, respectively, consistent with subcortical and cortical feedback latencies. Both LLRs increased with balance challenge and could be explained by center of mass kinematics. Perturbation-evoked antagonist muscle activity consisted of destabilizing and stabilizing components categorized based on whether they resist the kinematic errors that drive their activation. The destabilizing component occurred at similar to 180 ms and was negatively correlated with clinical measures of balance ability in the OA but not PD group. Exploratory comparisons showed OA and PD groups had larger LLR2s at lower challenge levels than YAs, consistent with greater cortical engagement during balance with aging. These findings demonstrate that a neuromechanical model can decompose perturbation-evoked muscle activity into hierarchical components related to clinical balance ability and identify mechanistic changes in the neural control of balance without direct brain measurements.
