Bioactive molecule design with geometric deep learning

Geometric deeplearning is a promising direction in molecular design and drug screening. Tomake sense of different representations and methods used in the field, a Reviewarticle by Kenneth Atz et al. provides anoverview of current principles and challenges. The cover image shows the resultof one such geometric deep learning approach, called DeepDock,developed by Oscar Méndez-Lucio and colleagues. The rat proteinPEPCK is shown as a 3D mesh in a binding conformation with potential smallmolecule drug 2-phosphoglycolate acid as predicted by the model. Theexperimentally validated conformation is superimposed in cyan.

See Kenneth Atz et al., and Méndez-Lucio et al.

A machine learning platform for the immune system

Deciphering the immune information encoded in adaptive immune receptor repertoires (AIRR) is important for next-generation immunodiagnostics and therapeutics design. However, the proliferation of molecular biology and bioinformatics tools that are necessary to generate large quantities of immune receptor data has not yet been matched by frameworks that allow for routine data analysis by biomedical scientists. Machine learning is a crucial technology for the transformation of AIRR into biomarkers of disease and infection, given its capacity to separate immunological signals from noise. In a paper in this issue, Pavlović et al. present immuneML, an open-source collaborative platform for AIRR machine learning. immuneML implements each step of the AIRR machine learning process using fully specified and shareable workflows and adapted to different user backgrounds.

See Pavlović et al.

Robotic body augmentation

Extra robotic limbs may enhance the physical abilities of healthy people or people with disabilities. However, complex applications will require humans to adapt and learn to operate a new robotic limb collaboratively with their biological limbs. In a paper in this issue, Dominijanni and colleagues describe the challenges that arise when controlling extra robotic limbs for body augmentation. They introduce the ‘neural resource allocation problem’, which is defined as the channelling of motor commands and sensory information to and from an augmentative device without hindering the motor control of biological limbs. The paper is highlighted in an editorial this issue.

See Dominijanni et al.

Crystal-structure phase mapping with deep reasoning networks

In scientific discovery, researchers don’t usually have large amounts of labelled data available and must solve complex problems using prior knowledge and reasoning to make sense of data. In a paper in this issue, Di Chen et al. https://doi.org/10.1038/s42256-021-00384-1 use a deep learning model called Deep Reasoning Network (DRNet) to tackle such a complex problem in materials science; the identification of crystal phases from the noisy mixture of X-ray diffraction signals. DRNets combine deep learning with constraint reasoning based on incorporated prior knowledge, in this case thermodynamic rules that govern crystal structures.

See Di Chen et al.

Quantized neural networks on the edge

As deep neural networks are pushed towards larger and more complex architectures, they require significant computational resources and are challenging to deploy in real-time applications. To reduce complexity, neural networks can be compressed, without a substantial decrease in model accuracy, by methods such as pruning or quantization. The latter involves using fewer bits to represent weights and biases. In a paper in this issue, Coelho Jr. et al. develop a method for producing quantized versions of deep neural network models for fully automated deployment on an FPGA chip with high-accuracy, low-energy and nanosecond inference, which is required for the real-time event-selection procedure in proton–proton collisions at the CERN Large Hadron Collider.

See Coelho Jr. et al.

Learning physical displacement fields with deep optical flow

Analysing complex flow dynamics is important in a wide range of problems in areas such as automotive, aerospace and biomedical engineering. Particle image velocimetry (PIV) is a key technique for visualizing and computing the velocity components of flow fields. Conventionally, manually designed algorithms are needed to process PIV measurements, but deep learning-based optical flow estimators are developed by Lagemann et al. that promise to be general, largely automated and to provide a high spatial resolution, which allows one to study very fine velocity fluctuations. The cover image highlights such a dense displacement field for a turbulent boundary layer predicted by the proposed optical flow learning model.

See Lagemann et al.

Coordinating safe navigation

The collective motion of swarms, such as flocks of birds or flying robots, can be described well in terms of locally defined rules, where every agent regulates its motion with respect to a limited set of neighbours. This complex phenomenon can be mathematically modelled using potential fields, but for aerial robots, this is not sufficient to guarantee safety in environments with obstacles. In this issue, Enrica Soria et al.. describe how a predictive model can incorporate the robots' dynamics and environments to improve the speed, order and safety of a swarm of aerial robots. The model is validated with a swarm of five quadrotors navigating a real-world indoor cluttered environment.

See Enrica Soria et al.

Neural architecture search for computational genomics

Applying deep learning models requires the tuning of network architectures for optimum performance, which can require substantial machine learning expertise. In this issue, Zijun Zhang et al. present a fully automated framework, AMBER, to design and apply convolutional neural networks for genomic sequences using neural architecture search. In an accompanying News & Views, Yi Zhang, Yang Liu and X. Shirley Liu discuss the AMBER technique and its potential to improve deep learning models in genomics.

See Zhang et al. and Zhang, Liu and Liu

Expanding dimensions

In the field of computational materials design, 3D microstructural datasets are crucial for understanding structure–performance relationships through physical modelling. However, 3D imaging can be slow and often has limited resolution compared to its 2D counterparts. In this issue, Steve Kench and Samuel Cooper propose a generative adversarial architecture, SliceGAN, which can use a single representative cross-sectional image to synthesize realistic 3D volumes. In an accompanying News & Views, Alejandro Franco discusses the technique and the potential to extend it to even further dimensional expansion.

See Kench and Cooper, and Franco

Deep learning for nanocrystal tomography

The 3D elemental structure and composition of nanocrystals can be analysed by combining scanning transmission electron microscopy (STEM) and energy-dispersive X-ray spectroscopy (EDX). This is useful, for instance, for the study and design of semiconductor quantum dots for optoelectronic applications in display devices. However, EDX has low efficiency and leads to electron beam-induced damage to the nanocrystals. In this issue, Han et al. demonstrate an unsupervised deep learning method that can help to reconstruct elemental 3D maps under reduced beam exposure. With this approach, valuable information can be learned about the dependence of optical properties on the structure and elemental composition of quantum dots.

See Han et al.

Tackling the torrent of scientific literature with active learning

Systematic reviews, which provide a comprehensive overview of the literature in a specific research area, are important tools for scholars, policymakers and clinicians, among others. However, producing them typically involves screening thousands (or tens of thousands) of papers, which is time-consuming and error-prone when done manually. This is frustrating especially when good overviews are urgently needed, for instance in the case of COVID-19 research. In recent years, it has become possible to speed up literature screening with machine learning approaches, and collaborative workflows have been developed where machine learning algorithms are optimized to find the most relevant records by human-in-the-loop approaches. In this issue van de Schoot et al. demonstrate an open source framework for machine learning-supported systematic reviewing of the literature called ASReview, integrating several machine learning classifiers.

See van de Schoot et al.

Learning visual appearance for flight control

Flying insects show impressive skills in navigation and piloting, including landing and avoiding obstacles, which roboticists try to mimic in the design of lightweight flying robots. The visual cue of optical flow is known to play a major role in insect navigation and is increasingly studied for use by small flying robots as well. However, there are gaps in the current understanding of optical flow control, as it cannot disentangle distance from velocity, and is less informative in the forward flight direction. In this issue, De Croon et al. propose a solution that consists of a learning process in which the robot first uses optical flow and self-induced oscillations to perceive distances to objects in its environment. It then learns a mapping from visual appearance to these distances to complement optical flow, solving the above-mentioned problems. The approach, which is biologically plausible in terms of processing, sensing, and actuation requirements, is demonstrated on a flying robot.

See De Croon et al.