Reviews & Analysis

We propose a computationally efficient genome-wide association study (GWAS) method, WtCoxG, for time-to-event (TTE) traits in the presence of case ascertainment— a form of oversampling bias. WtCoxG addresses case ascertainment bias by applying a weighted Cox proportional hazard model, and outperforms existing approaches when incorporating information on external allele frequencies.

A benchmark — MaCBench — is developed for evaluating the scientific knowledge of vision language models (VLMs). Evaluation of leading VLMs reveals that they excel at basic scientific tasks such as equipment identification, but struggle with spatial reasoning and multistep analysis — a limitation for autonomous scientific discovery.

An integrated platform, Digital Twin for Chemical Science (DTCS), is developed to connect first-principles theory with spectroscopic measurements through a bidirectional feedback loop. By predicting and refining chemical reaction mechanisms before, during and after experiments, DTCS enables the interpretation of spectra and supports real-time decision-making in chemical characterization.

We developed group technical effects (GTE) as a quantitative metric for evaluating gene-level batch effects in single-cell data. It identifies highly batch-sensitive genes — the primary contributors to batch effects — that vary across datasets, and whose removal effectively mitigates the batch effects.

Enhanced sampling methods aim to simulate rare physical and chemical reactive processes involving transitions between long-lived states. Existing methods often disproportionally sample either metastable or transition states. A machine-learning approach combines the strengths of these two cases to characterize entire rare events with the same thoroughness in a single calculation.

Predicting how molecular changes affect brain activity is a challenge in neuroscience. We introduced a multiscale modeling approach to simulate these microscopic changes and how they impact macroscale brain activity. This approach predicted how the anesthetic action on synaptic receptors can lead to the transitions in macroscale brain activity observed empirically.

Inspired by the morphologies of xeric plant leaves, we have developed biomimetic liquid crystal elastomer bilayers that can bend, spiral and twist. These adaptive shape morphing structures can twist to improve water collection efficiency and wind resistance, suggesting their potential application in adaptive water collection and directional transportation.

We introduce free-energy machine (FEM), an efficient and general method for solving combinatorial optimization problems. FEM combines free-energy minimization from statistical physics with gradient-based optimization techniques in machine learning and utilizes parallel computation, outperforming state-of-the-art algorithms and showcasing the synergy of merging statistical physics with machine learning.

We propose a diversity-aware population modeling framework using Bayesian multilevel regression and post-stratification to quantify sociodemographic disparities in cognitive development. Our approach improved subgroup estimates, guiding targeted public health strategies and addressing biases in traditional models to support more equitable decision-making.

Identifying pleiotropic associations for rare variants in multi-ethnic biobank-scale whole-genome sequencing data poses considerable challenges. This study introduced MultiSTAAR as a scalable and robust multi-trait rare variant analysis framework designed for both coding and noncoding regions by integrating multiple variant functional annotations and leveraging multivariate modeling across diverse phenotypes.

We present Spatial Modeling Algorithms for Reactions and Transport (SMART), a software package that simulates spatiotemporally detailed biochemical reaction networks within realistic cellular and subcellular geometries. This paper highlights the use of SMART in several biological test cases including cellular mechanotransduction, calcium signaling in neurons and cardiomyocytes, and adenosine triphosphate synthesis.

To achieve an advanced neuromorphic computing system with brain-like energy efficiency and generalization capabilities, we propose a hardware–software co-design of in-memory reservoir computing. This co-design integrates a liquid state machine-based encoder with artificial neural network projections on a hybrid analog–digital system, demonstrating zero-shot learning for multimodal event data.

We present Morpho, an extensible programmable environment that uses finite elements for shape optimization in soft matter. Given an energy functional that incorporates physical boundaries and effects such as elasticity and electromagnetism, together with additional constraints to be satisfied, Morpho predicts the optimized shape and structure adopted by the material.

We created an open-source model that simulates Caenorhabditis elegans in a closed-loop system, by integrating simulations of its brain, its physical body, and its environment. BAAIWorm replicated C. elegans locomotive behaviors, and synthetic perturbations of synaptic connections impacted neural control of movement and affected the embodied motor behavior.

Inspired by recent approaches for natural language processing and computer vision, we developed Annotatability, a framework that analyzes deep neural network training dynamics to interpret pre-annotated single-cell and spatial omics data. Annotatability identified erroneous annotations and ambiguous cell states, inferred trajectories from binary labels, and revealed underlying biological signals.

We developed mixture model inference with discrete-coupled autoencoders (MMIDAS), an unsupervised variational framework that jointly learns discrete clusters and continuous cluster-specific variability. When applied to unimodal or multimodal single-cell omic data, MMIDAS learned single-cell representations with robust cell type definitions and interpretable, continuous within-cell type variability.

Revealing a drug’s mechanism of action (MOA) is costly and time-consuming. In this study, we used deep learning to extract temporal mitochondrial phenotypic features after exposure to drugs with known MOAs using re-identification algorithms. The trained model could then predict the MOAs of unidentified substances, facilitating phenotypic screening-based drug discovery and repurposing.

A machine learning framework that uses atomic orbital-based Hamiltonian matrices and gradients predicted by an equivariant graph neural network is established to calculate electron–phonon coupling (EPC). This approach accelerates the calculations by several orders of magnitude, enabling EPC-related properties to be predicted for complex systems using highly accurate functionals.

Pretraining powerful deep learning models requires large, comprehensive training datasets, which are often unavailable for medical imaging. In response, the universal biomedical pretrained (UMedPT) foundational model was developed based on multiple small and medium-sized datasets. This model reduced the amount of data required to learn new target tasks by at least 50%.

A graph neural network using virtual nodes is proposed to predict the properties of complex materials with variable dimensions or dimensions that depend on the input. The method is used to accurately and quickly predict phonon dispersion relations in complex solids and alloys.