Articles in 2023

Optimal control of quantum many-body systems is needed to make use of quantum technologies, but is challenging due to the exponentially large dimension of the Hilbert space as a function of the number of qubits. Metz and Bukov propose a framework combining matrix product states and reinforcement learning that allows control of a larger number of interacting quantum particles than achievable with standard neural-network-based methods.

There are currently promising developments in deep learning for protein design, with applications in drug discovery and synthetic biology. For more efficient exploration of the design space, Wang et al. demonstrate a reinforcement learning method, EvoZero, for directed evolution in protein engineering towards desired functional or structure-related properties.

Out of the large number of neoepitopes, few elicit an immune response from the major histocompatibility complex. To predict which neoepitopes can be effective, Albert and colleagues present a method based on long short-term memory ensembles and transfer learning from immunogenicity assays.

Integrating gene expression across tissues is crucial for understanding coordinated biological mechanisms. Viñas et al. present a neural network for multi-tissue imputation of gene expression, exploiting the shared regulatory architecture of tissues.

Federated learning can be used to train medical AI models on sensitive personal data while preserving important privacy properties; however, the sensitive nature of the data makes it difficult to evaluate approaches reproducibly on real data. The MedPerf project presented by Karargyris et al. provides the tools and infrastructure to distribute models to healthcare facilities, such that they can be trained and evaluated in realistic settings.

Reaction–diffusion processes, which can be found in many fundamental spatiotemporal dynamical phenomena in chemistry, biology, geology, physics and ecology, can be modelled by partial differential equations (PDEs). Physics-informed deep learning approaches can accelerate the discovery of PDEs and Rao et al. improve interpretability and generalizability by strong encoding of the underlying physics structure in the neural network.

It is challenging to obtain a sufficient amount of high-quality annotated images for deep-learning applications in medical imaging, and practical methods often use a combination of labelled and unlabelled data. A dual-view framework builds on such semi-supervised approaches and uses two independently trained critic networks that learn from each other to generate segmentation masks in different medical imaging modalities.

This Reusability Report revisits a recently developed machine learning method for precision oncology, called ‘transfer of cell line response prediction’ (TCRP). Emily So et al. confirm the reproducibility of the previously reported results in drug-response prediction and also test the reusability of the method on new case studies with clinical relevance.

Computer-aided drug design has a high computational cost and a newly identified drug candidate might be unsuitable due to a range of drug properties. Lam and colleagues trained a model based on graph convolutional variational encoders that predicts a range of properties simultaneously to accelerate virtual screening.

Virtual worlds are typically encountered through simulated visual and auditory perceptions. Incorporating touch can create more immersive experiences with a sense of agency.

There are repeated calls in the AI community to prioritize data work — collecting, curating, analysing and otherwise considering the quality of data. But this is not practised as much as advocates would like, often because of a lack of institutional and cultural incentives. One way to encourage data work would be to reframe it as more technically rigorous, and thereby integrate it into more-valued lines of research such as model innovation.

The temporal nature of the transcriptome is important for understanding many biological processes, but it is challenging to measure. By leveraging datasets with multiple time series, Woicik and colleagues present a model that accurately extrapolates genomic measurements to unmeasured timepoints, including developmental gene expression, drug-induced perturbations and cancer gene mutations.

Mass spectrometry imaging (MSI) can provide important information, but long imaging times are needed to achieve a high spatial resolution, which is why the amount of publicly available high-resolution MSI data for deep learning applications is limited. Liao and colleagues use transfer learning from optical super-resolution images to reduce the amount of MSI data that is needed.

The stochastic features of memristors make them suitable for computation and probabilistic sampling; however, implementing these properties in hardware is extremely challenging. Lin et al. introduce an approach that leverages the cycle-to-cycle read variability of memristors as a physical random variable for in situ, real-time random number generation, and demonstrate it on a risk-sensitive reinforcement learning task.

A ‘programming’-like approach provides a one-step algorithm to find network parameters for recurrent neural networks that can model complex dynamical systems.

Sustainability awareness is lacking in the development of AI systems and algorithms for healthcare. The authors discuss resource sustainability issues in energy, storage and domain knowledge, and present potential solutions.

Recurrent neural networks are flexible architectures that can perform a variety of complex, time-dependent computations. Kim and Bassett introduce an alternative, ‘programming’-like computational framework to determine the appropriate network parameters for a specific task without the need for supervised training.

Robots that can change their shape offer flexible functionality. A modular robotic platform is shown that implements physical polygon meshing, by combining triangles with sides of adjustable lengths, allowing flexible three-dimensional shape configurations.