Research articles

A graph-based artificial intelligence model for urban planning outperforms human-designed plans in objective metrics, offering an efficient and adaptable collaborative workflow for future sustainable cities.

Real-world social networks are often ephemeral and subject to exogenous restructuring. Q. Su et al. show that dynamic networks can foster cooperative behavior.

Batch effects pose great statistical challenges to the analysis of biomedical data. One approach to address batch effects is through sample remeasurement in each batch. In this work, the researchers developed a rigorous batch-effect correction procedure based on remeasured samples.

Automating materials’ defect predictions with graph neural networks, when coupled to first principles thermodynamic calculations, accelerates materials discovery for a variety of high-temperature, clean-energy applications.

This study introduces a physically boosted cooperative learning framework (PBCL) to reveal 2D atmospheric turbulence strength fields from turbulence-distorted infrared images. The PBCL is further demonstrated to inhibit adverse turbulence effects in images.

A system called ORFanage can analyze RNA-seq data to find novel protein variants and improve gene annotations. In addition, the method is fast and scalable, being able to filter out noise and thus greatly improving the quality of transcriptome assemblies.

A Bayesian method, called MAGICAL, that contrasts single cell multiomics data across conditions to accurately discover differences in gene regulatory circuits at cell type resolution is applied to specific host-based diagnosis of bacterial sepsis.

MeDuSA, a mixed model-based method, leverages single-cell RNA-sequencing data for high-accuracy, fine-resolution cellular deconvolution in bulk RNA-sequencing data. It improves deconvolution accuracy over existing methods, revealing cell-state dynamics in various biological processes.

The study presents a mobility centrality index to delineate urban dynamics in quasi-real time with mobile-phone data. The results indicate that urban structures were becoming more monocentric during the COVID-19 lockdown periods in major cities in Spain.

Neural ADMIXTURE is a neural-network-based, interpretable autoencoder that performs rapid genomic clustering in biobank-scale databases.

GRAPE is a software resource for graph learning and embedding that is orders of magnitude faster than existing state-of-the-art libraries, making large-graph analysis feasible in a wide range of real-world applications.

A theoretical framework for quantum neural network (QNN) overparametrization, a phase transition in loss landscape complexity, is established. The precise characterization of the critical number of parameters offered is expected to impact QNN design.

A platform for single-cell meta-analysis of inflammatory bowel disease, named scIBD, enables identification of rare or less-characterized cell types and the dissection of the commonalities and differences between ulcerative colitis and Crohn’s disease.

This study presents an ab initio approach for the real-time charge carrier quantum dynamics in the momentum space, which is computationally more efficient than conventional real-space non-adiabatic molecular dynamics method. The method is applied to study hot carrier dynamics in graphene, which provides insights about the phonon-specific relaxation mechanism.

Kirigami is an ancient art form that is now increasingly studied and applied in science and technology. This work presents an additive approach for the computational design of kirigami and two fabrication strategies for its physical instantiation.

A disease space is constructed from clinical records by embedding all diseases and considering a patient’s space coordinates as a measure of their health state. This measure was associated with 108 genetic loci, on which models were built to predict various morbidities.

Mental conflict has been regarded as subjective instead of quantitative. This study developed a data-driven method to decode temporal dynamics of conflict between reward and curiosity, which can elucidate mechanisms of irrational decision-making.

This study proposes a diffusion model, ProteinSGM, for the design of novel protein folds. The designed proteins are diverse, experimentally stable and structurally consistent with predicted models

GAME-Net is a graph deep learning model trained with small molecules containing a wide set of functional groups for predicting the adsorption energy of closed-shell organic molecules on metal surfaces, avoiding expensive density functional theory simulations.

NeCLAS is a machine learning pipeline that can accurately and efficiently predict nanoscale interactions, which has broad applications in biological processes and material properties.