Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain
the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in
Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles
and JavaScript.
Earth system models (ESMs) are powerful tools for simulating climate fields, but weather forecasting and in particular precipitation prediction with ESMs are challenging. A generative adversarial network, constrained by the sum of global precipitation, is developed that substantially improves ESM predictions of spatial patterns and intermittency of daily precipitation.
Cell type annotation is a core task for single cell RNA-sequencing, but current bioinformatic tools struggle with some of the underlying challenges, including high dimensionality, data sparsity, batch effects and a lack of labels. In a self-supervised approach, a transformer model called scBERT is pretrained on millions of unlabelled public single cell RNA-seq data and then fine-tuned with a small number of labelled samples for cell annotation tasks.
The space of possible proteins is vast, and optimizing proteins for specific target properties computationally is an ongoing challenge, even with large amounts of data. Castro and colleagues combine a transformer-based model with regularized prediction heads to form a smooth and pseudoconvex latent space that allows for easier navigation and more efficient optimization of proteins.
The performance of machine learning models is usually compared via the mean value of a selected performance measure such as the area under the receiver operating characteristic curve on a specific benchmark data set. However, this measure, its mean value or even relative differences between models do not provide a good prediction of whether the results can translate to other data sets. Gosiewska and colleagues present here a comparison based on Elo ranking, which offers a probabilistic interpretation of how much better one model is than another.
Producing high-quality 3D refractive index maps from 2D intensity-only measurements is a long-standing objective in computational microscopy, with many applications involving the visualization of cellular and subcellular structures. A new method can reconstruct high-contrast and artefact-free images by employing the neural fields technique, which can learn a continuous 3D representation using a neural network that maps spatial coordinates to the refractive index values.
Genomic sequencing offers a wealth of information that could be analysed with deep neural networks. But despite good performance, neural networks can choose random features for their prediction. Kassani et al. present a method to stabilize the features selected by a DNN to make it more interpretable.
To understand reactions in organic chemistry, ideally simple rules would help us predict the outcome of new reactions, but in reality such rules are not easily identified. Chen and Jung extract generalized reaction templates from data and show that they can be used in graph neural networks to predict the outcome of reactions and, despite simplification, still represent a high percentage of existing reactions.
PROTACs can directly and selectively degrade proteins, which opens promising applications in the design of novel drugs, but designing effective PROTACs is a hard challenge due to the complexity of pharmacokinetics. Zheng et al. use a deep generative model to create likely candidates and screen them further to identify a novel BRD4-degrading PROTAC.
Changing weather conditions pose a challenge for autonomous vehicles. Almalioglu and colleagues use a geometry-aware learning technique that fuses visual, lidar and radar information, such that the benefits of each can be used under different weather conditions.
Kinetic models of metabolism capture time-dependent behaviour of cellular states and provide valuable insights into cellular physiology, but, due to the lack of experimental data, traditional kinetic modelling can be unreliable and computationally inefficient. A generative framework based on deep learning called REKINDLE can efficiently parameterize large-scale kinetic models, enabling new opportunities to study cellular metabolic behaviour.
Single-cell datasets continue to grow in size and complexity, calling for computational tools to process and analyse data. Yang et al. present a contrastive learning framework to learn cell representations from single-cell multiomics datasets.
Federated learning and unsupervised anomaly detection are common techniques in machine learning. The authors combine them, using multicentred datasets and various diseases, to automate the segmentation of brain abnormalities without the need for annotations or sharing private local data.
Finding stable radical compounds for redox flow batteries is a challenging molecular design task. Sowndarya et al. combine an AlphaZero-based framework with a surrogate objective function trained on quantum chemistry simulations to generate suitable radical candidates that are stable. The approach promises to contribute to the development of low-cost, reliable energy storage technologies.
Targeted drug delivery is an exciting application of nanorobotics, but directing particles in the blood stream to the right location and in sufficient number is challenging. Gu and colleagues have developed a microtubule scaffold with embedded micromagnets that allows cargo, such as drug particles, to be transported in microvascular networks with precision and speed.
So-called noisy intermediate-scale quantum devices will be capable of a range of quantum simulation tasks, provided that the effects of noise can be sufficiently reduced. A neural error mitigation approach is developed that uses neural networks to improve the estimates of ground states and ground-state observables of molecules and quantum systems obtained using quantum simulations on near-term devices.
Using the natural dynamics of a legged robot for locomotion is challenging and can be computationally complex. A newly designed quadruped robot called Morti uses a central pattern generator inside two feedback loops as an adaptive method so that it efficiently uses the passive elasticity of its legs and can learn to walk within 1 h.
Artificial DNA circuits that can perform neural network-like computations have been developed, but scaling up these circuits to recognize a large number of patterns is a challenging task. Xiong, Zhu and colleagues experimentally demonstrate a convolutional neural network algorithm using a synthetic DNA-based regulatory circuit in vitro and develop a freeze–thaw approach to reduce the computation time from hours to minutes, paving the way towards more powerful biomolecular classifiers.
An end-to-end machine learning approach that can learn which mechanisms determine cell fate and competition from a large time-lapse microscopy dataset is developed. The approach makes use of a probabilistic autoencoder to learn an interpretable representation of the organization of cells, and provides cell fate predictions that can be tested in drug screening experiments.
Deep learning methods can provide useful predictions for drug design, but their hyperparameters need to be carefully tweaked to give good performance on a specific problem or dataset. Li et al. present here a method that finds appropriate architectures and hyperparameters for a wide range of drug design tasks and can achieve good performance without human intervention.