Nature computational science最新文献

Current technologies enable the simultaneous measurement of diverse data types at the single-cell level. However, data are often processed separately, or integrated via representation learning methods that obscure the contributions of each data modality. Here we present a computational framework that automatically learns partial information sharing between multiple modalities by using an Autoencoder with a Partially Overlapping Latent space learned through Latent Optimization (APOLLO). We tested APOLLO on simulated data, and on four applications involving paired single-cell data: SHARE-seq (scRNA-seq and scATAC-seq), CITE-seq (scRNA-seq and protein abundance), and two multiplexed imaging datasets. APOLLO enables the prediction of missing modalities, such as unmeasured protein stains, and allows disentangling which modality or cellular compartment is linked with a specific phenotype, such as the variability in protein localization observed across single cells. Overall, APOLLO efficiently integrates diverse data modalities and, by retaining and distinguishing between shared and modality-specific information, provides a more interpretable and holistic view of cell state.

Raman hyperspectral imaging is a powerful technique for probing the intrinsic properties of samples by combining vibrational spectroscopy with spatial imaging. Despite its potential, the inherently weak Raman scattering signal typically necessitates prolonged acquisition times or high-power lasers, thereby limiting its efficiency and broader applicability. Here we present a computational method for facilitating Raman imaging under challenging conditions. We propose that even low-quality measurements-acquired with short integration times or low-power lasers-still contain sufficient information of Raman spectra. To this end, an unsupervised learning-based method, self-optimized spectral distance (SSD), is developed to reconstruct Raman images directly from 'noisy' measurements. By eliminating the dependence on large-scale training datasets, long imaging times and high-energy lasers, SSD helps to advance high-throughput Raman imaging. In diverse applications, including cellular structure analysis, microparticle detection and pharmaceutical ingredient identification, SSD achieves high imaging quality while reducing acquisition time and excitation power at least one order of magnitude.

Drawing on the experiences and lessons learned from researchers based in low- and middle-income countries (LMICs) that leverage generative artificial intelligence (GenAI) technologies to address socio-economic challenges, we showcase the considerable potential to use GenAI to accelerate the progress towards achieving some of the Sustainable Development Goals, as well as considerable obstacles for creating locally adapted AI tools for fair development in LMICs. An expanded evidence base on GenAI in resource-limited settings is crucial for policymakers to understand opportunities and risks, while rights-based safeguards against AI harms can be strengthened by the lived experiences of local projects.

A remarkable capability of the human brain is to form more abstract conceptual representations from sensorimotor experiences and flexibly apply them independent of direct sensory inputs. However, the computational mechanism underlying this ability remains poorly understood. Here we present a dual-module neural network framework, CATS Net, to bridge this gap. Our model consists of a concept-abstraction module that extracts low-dimensional conceptual representations, and a task-solving module that performs visual judgment tasks under the hierarchical gating control of the formed concepts. The system develops transferable semantic structure based on concept representations that enable cross-network knowledge transfer through conceptual communication. Model-brain fitting analyses reveal that these emergent concept spaces align with both neurocognitive semantic model and brain response structures in the human ventral occipitotemporal cortex, while the gating mechanisms mirror that in the semantic-control brain network. This work establishes a unified computational framework that can offer mechanistic insights for understanding human conceptual cognition and engineering artificial systems with human-like conceptual intelligence.

Salt-solvent chemistry underpins electrochemical systems, governing key properties such as ionic conductivity, viscosity and chemical stability. Yet, its rational design is hindered by the vast chemical space spanning countless combinations and nonlinear structure-behavior couplings, further amplified by sparse and imbalanced experimental data that impede generalization. Here we develop SCAN, a dynamic routing-guided framework for modeling and interpreting salt-solvent chemistry, which effectively handles long-tailed data and captures the full spectrum of salt-solvent formulations. We apply SCAN to non-aqueous electrolytes and achieve a benchmark error of 0.372 mS cm^-1 on conductivity, reducing predictive error by 65.3% over baselines. Then, we shape the conductivity atlas across 11,515,140 salt-solvent systems. Importantly, large-scale validations confirm a success rate of 81.08% for top-predicted candidates, including LiFSI-, LiTFSI- and LiBOB-based systems with conductivity >20 mS cm^-1. Beyond prediction, SCAN provides chemical insight into how molecular flexibility and ion-solvent interactions influence conductivity by incorporating the gradient-decoupling approach, symbolic regression and quantum chemistry calculation.

Here we introduce a graph neural network architecture built on geometric vector perceptrons to predict the committor function directly from atomic coordinates, bypassing the need for hand-crafted collective variables. The method offers atom-level interpretability, pinpointing the key atomic players in complex transitions without relying on prior assumptions. Applied across diverse molecular systems, the method accurately infers the committor function and highlights the importance of each heavy atom in the transition mechanism. It also yields precise estimates of the rate constants for the underlying processes. The proposed approach assists in understanding and modeling complex dynamics, by enabling collective-variable-free learning and automated identification of physically meaningful reaction coordinates of complex molecular processes.

The synthesis of crystalline materials, such as zeolites, remains a notable challenge owing to a high-dimensional synthesis space, intricate structure-synthesis relationships and time-consuming experiments. Here, considering the 'one-to-many' relationship between structure and synthesis, we propose DiffSyn, a generative diffusion model trained on over 23,000 synthesis recipes that span 50 years of literature. DiffSyn generates probable synthesis routes conditioned on a desired zeolite structure and an organic template. DiffSyn achieves state-of-the-art performance by capturing the multi-modal nature of structure-synthesis relationships. We apply DiffSyn to differentiate among competing phases and generate optimal synthesis routes. As a proof of concept, we synthesize a UFI material using DiffSyn-generated synthesis routes. These routes, rationalized by density functional theory binding energies, resulted in the successful synthesis of a UFI material with a high Si/Al_ICP of 19.0, which is expected to improve thermal stability.