Browse Articles
Discover research articles across all indexed journals
Thermodynamic Control of Facet Chemistry for Precise Solid-State Synthesis of Na Layered Cathodes
Ultra-wideband printed monopole antenna with cardioid-shaped fractal geometry
High-linearity all-2D heterojunction-based neuromorphic image-sensing transistor array with a sandwiched tunneling layer for dynamic vision
Interpretable deep learning framework for secure clinical data analytics
Abstract The growing use of artificial intelligence in the healthcare sector has facilitated the emergence of sophisticated clinical data analytics, but the most critical issues pertain to data privacy, model interpretability, and model trustworthiness. These problems still pose a challenge to its real-world application. This study proposes an interpretation of a deep learning framework for secure clinical data analytics. This proposes a privacy-preserving distributed learning framework in combination with a naturally interpretable model structure. The suggested framework enables parallel training across several healthcare facilities without compromising raw patient information, thereby keeping the data confidential without adversely affecting the analysis. An idea-driven deep learning framework is also presented to produce clinically significant intermediate representations that enable clear, interpretable predictions. Moreover, a new explanation stability mechanism is established to maintain the consistency and reliability of model explanations in distributed environments. To promote clinical safety, the decision module includes uncertainty to enable the system to detect predictions with low confidence and defer them to experts. Massive tests using actual clinical data on patient conditions can be conducted to demonstrate that the framework presented is much more successful than other methods with respect to predictiveness, interpretability, explanation stability, fairness, and calibration. Furthermore, the framework demonstrates that privacy preservation, interpretability, explanation reliability, and clinical usability can be jointly optimized within a unified learning architecture. The findings indicate the effectiveness of the suggested strategy in providing secure, transparent, and reliable clinical decision support, rendering it an appropriate strategy to implement in contemporary healthcare systems.
Fly neurons carry a molecular ‘timestamp’ that encodes their birth order
A unified network systems approach uncovers a core program underlying T follicular helper cell differentiation
Comparative analysis of machine learning and regularized regression models for predicting cardiovascular outcomes in patients with type 2 diabetes after acute coronary syndrome
Abstract Identifying type 2 diabetes (T2D) patients at high cardiovascular risk following acute coronary syndrome (ACS) remains challenging despite treatment advances. We systematically compared machine learning and regularized regression methods for predicting cardiovascular outcomes using data from the EXAMINE trial (NCT00968708). Prediction models incorporating over 100 clinical and biomarker variables were developed for three endpoints: cardiovascular death, nonfatal myocardial infarction, or stroke; all-cause mortality; and cardiovascular death or heart failure hospitalization. Eleven algorithms were evaluated using area under the ROC curve (AUC) at 12 and 24 months, with tenfold stratified cross-validation. Regularized regression methods showed the most consistent and generalizable performance, with substantial overlap in confidence intervals between top-performing methods. At 24 months, among regularized models LASSO achieved the highest testing AUC for cardiovascular death or heart failure hospitalization (0.80, 95% CI 0.75–0.85) and all-cause mortality (0.75, 95% CI 0.69–0.81), while the best regularized model for the primary composite endpoint reached an AUC of 0.69 (95% CI 0.64–0.74); the strongest discrimination overall was observed for the 12-month heart failure endpoint (AUC up to 0.83). Adding the proteomic panel to clinical predictors improved or maintained discrimination across all six outcome-by-horizon strata (mean ΔAUC + 0.03). Complex algorithms including neural networks and random forests exhibited overfitting despite cross-validation. These findings indicate that regularized regression methods provide the most consistent and generalizable performance for cardiovascular risk prediction in T2D patients with recent ACS. For the heart failure hospitalization endpoint, the leading regularized model outperformed the best-tuned tree-ensemble method (XGBoost) significantly at 12 months (ΔAUC + 0.065, 95% CI + 0.031 to + 0.100, DeLong p < 0.001) and by a consistent but non-significant margin at 24 months (ΔAUC + 0.037, 95% CI − 0.007 to + 0.082, p = 0.10). The superior performance for heart failure events suggests current clinical and proteomic markers are particularly valuable for heart failure risk stratification in this population.
AI-redesigned starting points and outcomes enhance protein evolution
Abstract Engineered or laboratory-evolved proteins often have suboptimal stability, activity or specificity. Here we applied artificial intelligence (AI)-based protein sequence design to address challenges in experimental enzyme evolution. Using the model ProteinMPNN, we redesigned three distinct botulinum neurotoxin (BoNT) proteases, generating variants with improved stability and full catalytic efficiency 1 . We hypothesized that redesigned enzymes may be more mutationally robust than their wild-type (WT) counterparts, and therefore may serve as better starting points to evolve new function. We performed side-by-side phage-assisted continuous evolution campaigns initiated with AI-redesigned proteases or with the corresponding WT proteases 2 . Evolving three distinct redesigned enzymes as starting points consistently yielded proteases with higher activity than evolving WT proteases in the same selection. Across four evolution campaigns, redesign conferred robustness that unlocked access to otherwise inaccessible highly functional sequences, confirmed by the inability of redesign-evolved mutations to function in WT enzyme backgrounds. When redesign raises fitness in sequence space local to the starting point, redesigned starting points adapt at a faster rate. Finally, we evolved both WT and AI-redesigned BoNT/E protease to selectively cleave the therapeutically relevant protein ataxin-2. Proteases evolved from the redesigned starting point reached higher catalytic efficiency and stability while minimizing native substrate cleavage, achieving more than 79-fold greater selected specificity for ataxin-2 than the best-performing variant evolved from WT BoNT/E. This study establishes a practical workflow using AI-redesigned starting points to evolve enzymes with improved properties compared with those evolved from natural proteins, with broad implications for protein science.
Cryo-EM insights into isoform-specific properties of the IP3R2 channel
An integrated framework of ChatGPT adoption in higher education using TAM and UTAUT models
Semiconducting and magnetic lanthanide MXenes from intercalated halides
Bariatric surgery resolves MASH by enhancing MAT1A-dependent one-carbon metabolism
A Programmable Cut-and-Sew Strategy for Trisulfide Transfer in the Synthesis of Cyclic Architectures
Superpixel based graph and random patch for polarimetric SAR image classification
The planktonic microbiome of the Great Barrier Reef
Abstract Large genome databases have markedly improved our understanding of marine microorganisms 1–5 . Although these resources have focused on prokaryotes, genomes from many dominant marine lineages, such as Pelagibacter and Prochlorococcus , are conspicuously underrepresented. Here we present the Great Barrier Reef Microbial Genomes Database (GBR-MGD), comprising 5,283 prokaryotic genomes obtained from Great Barrier Reef seawater samples using Nanopore and Illumina sequencing, including a collection of high-quality genomes of underrepresented groups. We show that standard short-read assemblies miss these populations owing to a combination of strain heterogeneity and low-GC-percentage sequencing bias. The GBR-MGD also comprises 20 chromosome-level picoeukaryote and 808,585 viral genomes, including a newly described clade of marine Crassvirales . We demonstrate the utility of the GBR-MGD to identify indicator taxa that can reliably predict the effects of reef management practices, such as the establishment of marine protected zones.
Molecular rotors reveal the 3D viscous habitat of mucus-colonizing bacteria
Abstract Interactions between bacteria and particulate organic matter, algae, coral reefs, fish, plant root systems, animals, and humans occur primarily through a dynamic interface of viscous mucus or mucilage. While mucus influences fundamental rates of bacterial infection, respiration, and carbon and nutrient cycling, our observations of this physical habitat of bacteria are limited by methods that damage the material and obfuscate spatial relationships. We present a technique using confocal microscopy of molecular rotors to reveal the 3D viscous structure of undisturbed mucus and associated bacteria. Quantification of the internal viscosity of mucus from different sources highlights variations in microscale morphologies that structure microbial distributions and ecological interactions. Individual examples of mucus aggregates from cultures of Chaetoceros affinis and Pseudo-nitzschia sp. diatoms exhibit consolidated versus patchy viscous morphologies, respectively, along with distinct patterns of microbial colonization. A viscous mucus layer surrounding an ephyra of the upside-down jellyfish Cassiopea xamachana may maintain a local microbiome while preventing direct contact with the animal. By quantifying the complex “mucoscape” shaping bacteria-organic matter interactions, this method provides a physical context for chemical fluxes and microbial activity in diverse ecosystems.