Research articles

Zeroth-order optimization is used on problems where no explicit gradient function is accessible, but single points can be queried. Hoffman et al. present here a molecular design method that uses zeroth-order optimization to deal with the discreteness of molecule sequences and to incorporate external guidance from property evaluations and design constraints.

The black-box nature of neural networks is a concern for high-stakes medical applications in which decisions must be based on medically relevant features. The authors develop an interpretable machine learning-based framework that aims to follow the reasoning processes of radiologists in providing predictions for cancer diagnosis in mammography.

The COVID-19 pandemic sparked the need for international collaboration in using clinical data for rapid development of diagnosis and treatment methods. But the sensitive nature of medical data requires special care and ideally potentially sensitive data would not leave the organization which collected it. Xiang Bai and colleagues present a privacy-preserving AI framework for CT-based COVID-19 diagnosis and demonstrate it on data from 23 hospitals in China and the United Kingdom.

To train deep learning methods to segment very small subcellular structures, the training data have to be labelled by experts as the optical effects at such a small scale and the narrow depth of focus make it difficult to identify individual structures. Sekh et al. use a physics-based simulation approach to train neural networks to automatically segment subcellular structures despite the optical artefacts.

Predicting the function of proteins in newly sequenced organisms is a challenging problem. Mateo Torres et al. present here a method to transfer the functional relations from known organisms and improve the prediction using network diffusion.

To improve desired properties of drugs or other molecules, deep learning can be used to guide the optimization process. Chen et al. present a method that optimizes molecules one fragment at a time and requires fewer parameters and training data while still improving optimization performance.

Predicting binding of ligands to molecular targets is a key task in the development of new drugs. To improve the speed and accuracy of this prediction, Méndez–Lucio and colleagues developed DeepDock, a method that uses geometric deep learning to inform a statistical potential to find conformations of ligand–target pairs.

Radiofrequency pulses of different shapes can increase the efficiency of applications such as broadcasting or medical imaging, but finding the optimal shape for a specific use can be computationally costly. Shin and colleagues present a new method based on deep reinforcement learning to design radiofrequency pulses for use in MRI, which is demonstrated to cover different types of optimization goals for each application.

RNA structure profiling methods suffer from missing values in RNA structurome data. Inspired by a computer vision approach, Gong and colleagues develop a deep learning method that imputes missing RNA structure scores and increases the structural coverage of the transcriptome.

The proliferation of molecular biology and bioinformatics tools necessary to generate huge quantities of immune receptor data has not been matched by frameworks that allow easy data analysis. The authors present immuneML, an open-source collaborative ecosystem for machine learning analysis of adaptive immune receptor repertoires.

Identifying a chemical substance using mass spectrometry without knowing its structure is challenging. To help detect novel designer drugs from their mass spectra, Skinnider et al. describe a generative model that is biased towards creating potentially psychoactive molecules and thus helps identify potential candidates for a specific sample.

Providing patient specific predictions for drug responses is challenging as preclinical data across a large population is hard to collect. Sharifi-Noghabi and colleagues present a semi-supervised method to predict drug response from limited data that can generalize successfully to different tissue types.

Complex physical processes such as flow fields can be predicted using deep learning methods if good quality sensor data is available, but sparsely placed sensors and sensors that change their position present a problem. A new approach from Kai Fukami and colleagues based on Voronoi tessellation now allows to use data from an arbitrary number of moving sensors to reconstruct a global field.

Optimization problems can be described in terms of a statistical physics framework. This offers the possibility to make use of ‘simulated annealing’, which is a procedure to search for a target solution similar to the gradual cooling of a condensed matter system to its ground state. The approach can now be sped up significantly by implementing a model of recurrent neural networks, in a new strategy called variational neural annealing.

Camera trapping is a widely adopted method for monitoring terrestrial mammals. However, a drawback is the amount of human annotation needed to keep pace with continuous data collection. The authors developed a hybrid system of machine learning and humans in the loop, which minimizes annotation load and improves efficiency.

Combining generative models and reinforcement learning has become a promising direction for computational drug design, but it is challenging to train an efficient model that produces candidate molecules with high diversity. Jike Wang and colleagues present a method, using knowledge distillation, to condense a conditional transformer model to make it usable in reinforcement learning while still generating diverse molecules that optimize multiple molecular properties.

A growing number of researchers are developing approaches to improve fairness in machine learning applications in areas such as healthcare, employment and social services, to avoid propagating and amplifying racial and other inequities. An empirical study explores the trade-off between increasing fairness and model accuracy across several social policy areas and finds that this trade-off is negligible in practice.

Turbulent optical distortions in the atmosphere limit the ability of optical technologies such as laser communication and long-distance environmental monitoring. A new method using adversarial networks learns to counter the physical processes underlying the turbulence so that complex optical scenes can be reconstructed.

The use of sparse signals in spiking neural networks, modelled on biological neurons, offers in principle a highly efficient approach for artificial neural networks when implemented on neuromorphic hardware, but new training approaches are needed to improve performance. Using a new type of activity-regularizing surrogate gradient for backpropagation combined with recurrent networks of tunable and adaptive spiking neurons, state-of-the-art performance for spiking neural networks is demonstrated on benchmarks in the time domain.