Accurate predictive thermodynamic models are essential tools for the computational design of new molecules. Models based on COnductor-like Screening Models (COSMO), such as COSMO-SAC and COSMO-RS, are well-suited for this purpose but require computationally expensive quantum-mechanical (QM) calculations to generate key properties - the surface charge density distribution (p(σ)), the surface area (A) and the cavity volume (V) - for each molecule. Here we develop graph-based neural network surrogate models, based on a directed message passing neural network (DMPNN) and a graph convolutional network (GCN), to predict these molecular properties rapidly and accurately. We train the models on a dataset of over 16 000 compounds generated with an automated QM calculation pipeline. The DMPNN model outperforms GCN for the prediction of the surface charge density, while the GCN provides a higher accuracy for surface area and cavity volume. We build on these strengths to propose a hybrid model: COSMO-NET. When applied to predict octanol-water partition coefficients and solid-liquid solubility across diverse solvents, COSMO-NET consistently outperforms individual models with the ranking: COSMO-NET > DMPNN > GCN > ECFP-MLP. COSMO-NET demonstrates remarkable robustness with minimal performance variation across different data splitting strategies. These results demonstrate that machine-learning surrogates can reliably replace costly QM calculations while maintaining accuracy, enabling the rapid discovery of new molecules and accurate evaluation of their thermodynamic properties.

Retinal Diseases are the main reason for vision loss and blindness worldwide. Initial identification of retinal abnormalities is essential for avoiding severe retinal-related complications and improving patient results. Traditional retinal disease diagnosis mainly depends on manual checks of retinal fundus images, which are slow and highly dependent on medical expertise. Deep learning methods are used in retinal disease classification. Existing methods are primarily based on labelled image data and often do not fully utilize complementary clinical information. The proposed hybrid multi-model framework classifies retinal diseases from the available retinal fundus images.

Brown adipose tissue (BAT) is a metabolically active tissue in humans, located primarily within the supraclavicular adipose tissue (scAT), that can be activated by cold or high-caloric meal consumption. While the changes of proton density fat fraction (PDFF) upon cold activation are well investigated, there is a knowledge gap regarding PDFF and T 2 ∗ dynamics upon meal-induced BAT activation. Therefore, the purpose of this study was to establish a robust pipeline to quantify PDFF and T 2 ∗ changes in scAT in fasting control and meal intervention experiments. To that end, a motion-robust radial stack-of-stars trajectory has been employed with a regularized iterative reconstruction algorithm, graph-cut-based water-fat separation, affine image registration and automatic neural-network-based region of interest (ROI) segmentation. Next to scAT, ROIs of subcutaneous adipose tissue (SAT) and muscle have been created to compare the results to reference tissues. Multi-echo gradient echo acquisitions were repeated in a total of ten volunteers before a break, in which the volunteers consumed an iso-caloric meal in the meal intervention experiment. After the break, repeated multi-echo gradient echo measurements were used to dynamically track the changes of both PDFF and T 2 ∗ . In the meal intervention experiment, scAT PDFF showed a stronger decrease of ∼ 1 % PDFF compared to the fasting control experiment, in which the decrease below baseline was similar to that in SAT. Furthermore, scAT T 2 ∗ exhibited a strong decrease to - 3 . 2 ms below baseline after meal intervention, while no decrease was observed in the fasting control experiment or in SAT. Considering the observed dynamic effects and their differences to the fasting control experiment, a PDFF decrease of ∼ 1 % PDFF and a T 2 ∗ decrease of ∼ 3 ms could be attributed to meal-induced activation of scAT.

High rates of comorbid posttraumatic stress disorder (PTSD) and alcohol use disorder (AUD) may stem from a shared neurobiological basis; however, research exploring this overlap remains limited. We examined whole-brain structural covariance (SC) across PTSD, AUD, comorbid PTSD & AUD, trauma-exposed controls, and trauma-naïve controls to identify shared and unique SC profiles. Structural brain scans were acquired from individuals across 14 sites (N = 1565), and cortical thickness estimates were extracted to compute SC. Group differences in SC between pairs of cortical regions were examined using the Network-Based statistic and quantified using local and global graph theory (GT) measures. Clinical groups showed consistent alterations in SC compared with trauma-naïve controls, characterised by lower SC involving frontal regions, particularly the inferior frontal gyrus and frontal pole, which also demonstrated high centrality in local GT analyses. In contrast, higher SC was observed in temporal regions, with the fusiform gyrus consistently emerging as a central region. Comparisons involving PTSD & AUD also showed increased SC in occipital regions, suggesting a potentially distinct pattern of structural organization in this group. Global GT analyses demonstrated altered network organization in clinical groups, including increased clustering and path length, and reduced global efficiency relative to trauma-naïve controls. Local efficiency was increased in PTSD and PTSD & AUD, but not AUD alone. Together, these findings indicate that alterations in frontal and temporal regions are implicated across PTSD, AUD, and their comorbidity. However, the PTSD & AUD group also exhibited additional occipital lobe alterations, suggesting a broader distribution of structural covariance alterations in the comorbid group.

Accurate identification of vehicle-group-level traffic risk is important for intelligent transportation safety management. Existing risk-prediction studies have mainly focused on individual vehicles, pairwise interactions, or aggregated surrogate safety indicators, while the role of group-level structure-behavior coupling remains insufficiently examined. To address this issue, this paper proposes a leakage-aware structure-behavior graph learning framework for Vehicle Group (VG) risk modeling. First, time-resolved VG graphs are constructed from high-frequency MAGIC trajectory data using an impact-induced grouping strategy with controlled supplementary spatial adjacency. Second, node-level surrogate risk states are defined using inverse Time-to-Collision (iTTC) and activated Post-Encroachment Time (PET), and then aggregated into VG-level surrogate risk labels through a group-risk ratio. Third, structural descriptors and behavioral features are integrated within a Graph Attention Network (GAT) for VG-level risk classification. To improve methodological transparency, the revised framework explicitly justifies the surrogate-labeling thresholds using empirical distributions, threshold tradeoff curves, and sensitivity analyses. It also includes leakage-audit experiments to examine whether model performance is driven by label-proximal surrogate quantities. Experiments on the MAGIC dataset show that the proposed model achieves strong and stable performance under repeated random seeds, with F1-score = 0.960 ± 0.002, ROC-AUC = 0.949 ± 0.008, PR-AUC = 0.991 ± 0.001, and Brier score = 0.058 ± 0.002. Compared with linear, neural, tree-based, and graph-based baselines, the proposed model provides competitive risk-identification performance, particularly in recall, false-negative control, and positive-class ranking. Additional robustness, prospective-prediction, and interpretability analyses indicate that the learned structure-behavior representation remains informative under temporal and sampling-frequency shifts and can highlight risk-relevant vehicles within VGs. The findings should be interpreted as evidence of structure-aware predictive association under a surrogate-labeling framework, rather than as proof that VG risk is inherently structural or that the model is deployment-ready. Overall, this study provides a more transparent and auditable graph-learning approach for VG-level surrogate risk modeling, offering decision-support insights for future CAV-oriented traffic safety management subject to external validation and deployment-oriented computational testing.

Sphingolipid dysregulation is an emerging feature of amyotrophic lateral sclerosis (ALS), but no systematic method separates species that may worsen axon-myelin membrane integrity from matter that may counteract it. We present an in-silico framework classifying sphingolipids and designed derivatives as candidate membrane-disruptive or candidate protective species, scored against real biological referents using externally published tools rather than a bespoke composite. Its central finding concerns a single, chemically defined difference in lipid state. In coarse-grained (MARTINI 3) molecular dynamics with three independent replicates per system, intact palmitoyl-sphingomyelin is indistinguishable from a pure-POPC bilayer (leaflet-mean Δ thickness + 0.001 nm, 95% CI brackets zero), whereas the modeled N-deacylated form thins it in every replicate (Δ=-0.043 nm, [- 0.046, - 0.041]), tracking a lyso-phosphatidylcholine detergent control (-0.083 nm, [- 0.087, - 0.080]) on two concordant estimators. Because only one lyso-sphingolipid is parameterized in public MARTINI 3, N-deacylation is not established here as a class-wide effect. Across eleven ALS studies (2021-2026), six report elevated serine-palmitoyltransferase (SPT) pathway sphingolipids, all in monogenic SPTLC1/SPTLC2 cohorts, whereas sporadic cohorts show disturbance without a shared direction. Because the simulations do not address whether this species is produced, accumulates, or reaches pathological concentrations in ALS tissue, we present the ALS connection as hypothesis-generating rather than direct evidence. Separately, an explicitly exploratory arm reports two designed modulators with predicted target engagement (Boltz-2) and an ADMET triage of their liabilities. The primary contribution is the membrane result with a wet-lab panel that makes it falsifiable.

Dieticyclidine (DTC) is a psychoactive compound for which efficient capture and identification remain analytically relevant. In this study, density functional theory calculations were used to comparatively screen pristine C60 and the substitutionally doped BC59 and SiC59 nanocages as candidate platforms for DTC adsorption. Gas-phase calculations were performed at the B97-D/6-31G(d) level, whereas aqueous-phase behavior was examined using M06-2X/LANL2DZ with a polarizable continuum model. Optimized geometries, molecular electrostatic potentials, frontier orbitals, density-of-states profiles, dipole moments, BSSE-corrected adsorption energies, recovery times, electronic-response descriptors, natural bond orbital interactions, and ELF/LOL maps were analyzed. Doping generated localized interaction sites and shortened the adsorbate-nanocage contacts relative to pristine C60. The calculated adsorption energies were -2.98, -19.35, and -29.87 kcal mol[-1] in the gas phase and -4.93, -24.57, and -41.31 kcal mol[-1] in water for C60@DTC, BC59@DTC, and SiC59@DTC, respectively. Adsorption also reduced the HOMO-LUMO gaps and produced positive DTC-fragment charges, suggesting electron-density transfer from DTC to the nanocages. NBO and real-space localization analyses further indicated that B and Si substitution strengthens lone-pair-assisted, polarized donor-acceptor interactions with possible partial bonding character, while not providing definitive evidence for conventional covalent B-N or Si-N bond formation. Overall, the Eads-based recovery descriptor qualitatively indicated lower relative retention for C60@DTC, intermediate retention for BC59@DTC, and the highest relative retention for SiC59@DTC. These values were used only for comparative screening and were not interpreted as quantitative desorption times. These findings constitute a pre-synthesis screening framework and require experimental validation under realistic conditions.

Direct air capture (DAC) requires sorbents capable of binding CO2 at a partial pressure of 0.0004 bar, where conventional flue-gas screening criteria may not define absorption performance. Machine-learning (ML) surrogates for Grand Canonical Monte Carlo (GCMC) simulation can accelerate screening of metal-organic frameworks (MOFs) for DAC applications. However, the relative performance of compositional featurization approaches under DAC conditions remains insufficiently explored. In this study, CO2 uptake at 0.0004 bar and 298 K was derived for 2523 experimentally realized CoRE-MOF-2019 frameworks by Langmuir fitting of NIST GCMC isotherms yielding a mean fit R[2] of 0.987. Two ML pipelines were developed using the same data and splits, and combined CoRE-MOF geometric descriptors with either 93 Matminer compositional features or 53 JARVIS-CFID chemical-interaction features. Optuna-optimized gradient-boosting regressor and LightGBM models achieved test-set R[2] of 0.82 and 0.83 for Matminer and JARVIS-CFID feature spaces, respectively. External validation on hold-out sets comprising 839 frameworks yielded R[2] values of 0.76-0.81 with Pearson correlation coefficients above 0.92. Feature-importance, SHAP, and LIME analyses consistently showed that volumetric surface area, Langmuir saturation capacity, and pore dimensions carry more than 90% of model attribution, while compositional descriptors contribute an additional contribution of approximately 7%. Partial-dependence analysis revealed a largest cavity diameter of 7-8 Å and pore-limiting diameter of 5 Å, combined with volumetric surface area above 1500 m[2]cm[-3], as key thresholds associated with higher CO2 uptake. These results indicate that pore geometry is dominant in determinant of CO2 uptake under DAC conditions, while chemical composition provides secondary effect.

Accurate identification of aortic dissection (AD) in emergencies is clinically critical. Non-contrast CT (NC-CT), the standard emergency imaging for chest pain, has limited AD diagnostic sensitivity. We developed an anatomy prior-driven multimodal contrastive learning framework, extracting discriminative features from contrast-enhanced CT (CE-CT) and radiological reports to enhance NC-CT-based AD detection. Aortic segmentation and straightening focus analysis on relevant anatomy, while our tri-encoder architecture leverages Fourier cross-attention (core innovation) to co-extract spatial and frequency-domain features with hybrid contrastive-focal frequency loss. After pre-training, only NC-CT is needed for inference. Evaluated on 134 subjects (70 AD, 64 non-AD), our method achieved 0.955 accuracy and 0.983 AUC, outperforming 9 SOTA models. External validation showed 0.933 accuracy, with lumen segmentation reaching an average Dice score of 0.773. These results demonstrate that our method can advance AD detection and enable accurate true/false lumen localization using NC-CT alone.