Research Briefing

SANGO efficiently removed batch effects between the query and reference single-cell ATAC signals through the underlying genome sequences, to enable cell type assignment according to the reference data. The method achieved superior performance on diverse datasets and could detect unknown tumor cells, providing valuable functional biological signals.

Approaches are needed to accelerate the discovery of transition metal complexes (TMCs), which is challenging owing to their vast chemical space. A large dataset of diverse ligands is now introduced and leveraged in a multiobjective genetic algorithm that enables the efficient optimization of TMCs in chemical spaces containing billions of them.

We present SCORPION, a computational tool to model gene regulatory networks based on single-cell transcriptomic data and prior knowledge of gene regulation. SCORPION networks can be modeled for specific cell types in individual samples, and are therefore suitable for conducting comparisons between experimental groups.

The laws of physics, formulated in a compact form, are elusive for complex dynamic phenomena. However, it is now shown that, using artificial intelligence constrained by the physical Onsager principle, a custom thermodynamic description of a complex system can be constructed from the observation of its dynamical behavior.

Transformer methods are revolutionizing how computers process human language. Exploiting the structural similarity between human lives, seen as sequences of events, and natural-language sentences, a transformer method — dubbed life2vec — has been used to create rich vector representations of human lives, from which accurate predictions can be made.

Using deep learning methods to study gene regulation has become popular, but designing accessible and customizable software for this purpose remains a challenge. This work introduces a computational toolkit called EUGENe that facilitates the development of end-to-end deep learning workflows in regulatory genomics.

A pairwise binding comparison network (PBCNet) has been established for predicting the relative binding affinity among congeneric ligands, using a physics-informed graph attention mechanism with a pair of protein pocket-ligand complexes as input. PBCNet shows practical value in guiding structure-based drug lead optimization with speed, precision, and ease-of-use.

Inspired by the classic lock-and-key model and advances in equivariant deep network design, we present a structure-based drug design model, SurfGen, which uses two types of equivariant graph neural networks to learn on protein surfaces and geometric structures to directly design small-molecule drugs.

We introduce STAligner — a graph neural network-based tool for the integration of multiple spatial transcriptomics datasets by generating batch effect-corrected embeddings, thereby enabling consensus spatial domain identification and accurate 3D tissue reconstruction.

Different cells can have very different three-dimensional morphologies. We present the computational framework u-signal3D that calculates the spatial scales at which molecules are organized on the surfaces of heterogeneously shaped cells, enabling high-throughput analyses and subsequent machine learning applications.

Graph neural networks (GNNs) present a promising route for machine learning of solid-state materials’ properties, but methods capable of directly predicting defect properties from ideal, defect-free structures are needed. A GNN developed for direct defect property predictions enables high-throughput screening of redox-active oxides for energy applications and beyond.

The dissipation and bending of light waves by atmospheric turbulence adversely affects infrared imaging, leading to grayscale drift, distortion, and blurring. A deep learning method has been developed to both extract the two-dimensional atmospheric turbulence strength fields and obtain clear and stable images from turbulence-distorted infrared images.

Increasing the number of parameters in a quantum neural network leads to a computational ‘phase transition’, beyond which training the network becomes significantly easier. An algebraic theory has been developed for this overparametrization phenomenon and predicts its onset above a certain parameter threshold.

GRAPE is a software resource for graph processing, learning and embedding that is orders of magnitude faster than existing state-of-the-art libraries. GRAPE can quickly process real-world graphs with millions of nodes and billions of edges, enabling complex graph analyses and research in graph-based machine learning and in diverse disciplines.

By conducting single-cell meta-analyses of inflammatory bowel disease, we identify rare or less-characterized cell subtypes linked to GWAS risk genes and therapeutic targets and dissect the commonalities and differences between ulcerative colitis and Crohn’s disease. Consequently, we present an interactive and user-friendly platform for the research community.

We often encounter mental conflict in our lives. Such mental conflict has long been regarded as subjective. However, a machine learning method can be used to quantify the temporal dynamics of conflict between reward and curiosity from behavioral time-series.

A graph neural network — GAME-Net — has been developed to predict the adsorption energy of organic molecules on metal surfaces, which is a key descriptor of heterogeneous catalytic activity. This method allows for the study of large molecules derived from raw materials such as plastic waste, avoiding the use of costly and time-intensive first-principles simulations.

A rotational and time-reversal equivariant neural network designed to represent the spin–orbital density functional theory Hamiltonian as a function of the atomic and magnetic structure enables ab initio electronic-structure calculations of magnetic superstructures. These calculations can efficiently and accurately predict subtle magnetic effects in various chemical environments.

A machine learning algorithm has been developed to capture and analyze rare molecular processes, revealing how molecules self-organize and function. The algorithm is general and can be applied whenever a dynamic system has a notion of ‘likely fate’.

This work involved the design of a multi-view manifold learning algorithm that capitalizes on various types of structure in high-dimensional time-series data to model dynamic signals in low dimensions. The resulting embeddings of human functional brain imaging data unveil trajectories through brain states that predict cognitive processing during diverse experimental tasks.