Articles

The study shows that scale-specific oscillations and scale-free neuronal avalanches in resting brains co-exist in the simplest model of an adaptive neural network close to a non-equilibrium critical point at the onset of self-sustained oscillations.

A biasing energy derived from the uncertainty of a neural network ensemble modifies the potential energy surface in molecular dynamics simulations to rapidly discover under-represented structural regions that meaningfully augment the training data set.

A topological data analysis-driven machine learning model for guiding protein engineering is proposed, complementing protein sequence and structure embeddings when navigating the fitness landscape.

The concept of evolving scattering networks is proposed for material design in wave physics. The concept has the potential to enable network-based material classification, microstructure screening and the design of stealthy hyperuniformity with superdense phases.

A generative deep learning model of molecular structure is combined with supervised deep learning models of molecular properties to achieve high-throughput (multi-)property-driven design of organic molecules.

A method that infers gene networks and rate parameters directly from single-molecule fluorescence in situ hybridization RNA snapshot data is proposed and demonstrated on synthetic and real data, providing insights on data from S. cerevisiae and E. coli.

A framework is presented to extrapolate the range of behaviors for influenza antibodies. Using this basis set of behaviors, the collective action of multiple antibodies can be teased apart to describe the individual antibodies within.

Design choices for dimensionality reduction on calcium imaging recordings are systematically evaluated, and a method called calcium imaging linear dynamical system (CILDS) is proposed for performing deconvolution and dimensionality reduction jointly.

This work provides a physics-based theoretical framework for accurate protein–ligand binding affinity estimation based on molecular dynamics simulations, enhanced sampling, non-parametric reweighting and the orientation quaternion formalism.

A hybrid functional (CF22D) with higher across-the-board accuracy for chemistry than most existing non-doubly hybrid functionals is presented by using a large database and a performance-triggered iterative supervised training method.

A density functional recommender enables chemical space exploration by selecting the best exchange–correlation functional for each system, outperforming the use of a single functional for all systems or transfer learning models.

In this study, a supervised protein language model is proposed to predict protein structure from a single sequence. It achieves state-of-the-art accuracy on orphan proteins and is competitive with other methods on human-designed proteins.

This study presents a model-agnostic framework that pairs deep neural operators and Bayesian experimental design for the accurate prediction of extreme events, such as rogue waves, pandemic spikes and structural ship failures.

Partial differential equations are typically solved on every element of a discretization basis before extracting the desired information, and each input requires one solution. In this study, a strategy is proposed to directly compute the quantities of interest, bypassing full-basis solutions and avoiding repetition over inputs.

Three machine learning methods are developed for discovering physically meaningful dimensionless groups and scaling parameters from data, with the Buckingham Pi theorem as a constraint.

The indeterminacy of edge or surface energy — unknowable for low-symmetry crystals — is avoided by an algebraic system complemented with closure equations, which enables computing algorithms to predict the equilibrium shape of any crystal.

A neural network-based language model of supra-word meaning, that is, the combined meaning of words in a sentence, is proposed. Analysis of functional magnetic resonance imaging and magnetoencephalography data helps identify the regions of the brain responsible for understanding this meaning.

Energy gradients from accurate post-density-functional-theory methods often involve statistical uncertainties. This study proposes an algorithm for structural optimizations when using gradients originating from such methods.

A machine learning interatomic potential model is designed and trained on diverse crystal data comprising 89 elements, enabling materials discovery across a vast chemical space without retraining.

A method is proposed to predict drug-induced single-cell gene-expression profiles with tensor imputation, detect cell type-specific maker genes, and identify the trajectories of regulated pathways considering intercellular heterogeneity.