Articles in 2021

A protein’s three-dimensional structure and properties are defined by its amino-acid sequence, but mapping protein sequence to protein function is a computationally highly intensive task. A new generative adversarial network approach learns from natural protein sequences and generates new, diverse protein sequence variations, which are experimentally tested.

While deep learning models have allowed the extraction of fingerprints from the structural description of molecules, they can miss information that is present in the molecular descriptors that chemists use. Shen and colleagues present a method to combine both sources of information into two-dimensional fingerprint maps, which can be used in a wide variety of pharmaceutical tasks to predict the properties of drugs.

Several technology companies offer platforms for users without coding experience to develop deep learning algorithms. This Analysis compares the performance of six ‘code-free deep learning’ platforms (from Amazon, Apple, Clarifai, Google, MedicMind and Microsoft) in creating medical image classification models.

Computational models that capture the nonlinear processing of the inner ear have been prohibitively slow to use for most machine-hearing systems. A convolutional neural network model replicates hallmark features of cochlear signal processing, potentially enabling real-time applications.

The propagation of ultrashort pulses in optical fibres, of interest in scientific studies of nonlinear systems, depends sensitively on both the input pulse and the fibre characteristics and normally requires extensive numerical simulations. A new approach based on a recurrent neural network can predict complex nonlinear propagation in optical fibre, solely from the input pulse intensity profile, and helps to design experiments in pulse compression and ultra-broadband supercontinuum generation.

Hearing and vision are powerful and important senses for interacting with our surroundings. So far, advances in the area of machine vision have been the most prominent, but machine hearing research that closely mimics the complex sound processing in the human ear has exciting opportunities to offer.

Many researchers have become interested in implementing artificial intelligence methods in applications with socially beneficial outcomes. To provide a way to study and benchmark such ‘AI for social good’ applications, Josh Cowls et al. use the United Nations’ Sustainable Development Goals to systematically analyse AI for social good applications.

The Conference on Neural Information Processing Systems (NeurIPS) introduced a new requirement in 2020 that submitting authors must include a statement on the broader impacts of their research. Prunkl and colleagues discuss challenges and benefits of this requirement and propose suggestions to address the challenges.

Chemical reactions can be grouped into classes, but determining what class a specific reaction belongs to is not trivial on a large-scale. A new study demonstrates data-driven automatic classification of chemical reactions with methods borrowed from natural language processing.

Extensive work has gone into developing peripheral auditory models that capture the nonlinear processing of the ear. But the resulting models are prohibitively slow to use at scale for most machine hearing systems. The authors present a convolutional neural network model that replicates hallmark features of cochlear signal processing, potentially enabling real-time applications.

Advanced electron microscopy and spectroscopy techniques can reveal useful structural and chemical details at the nanoscale. An unsupervised deep learning approach helps to reconstruct 3D images and observe the relationship between optical and structural properties of semiconductor nanocrystals, of interest in optoelectronic applications.

Model interpretability is important in genomics. Koo and Ploenzke show that exponential activations in the first layer of convolutional neural networks lead to interpretable and robust representations of genomic sequence motifs.

Cryo-electron microscopy (cryo-EM) can be used to determine the three-dimensional structure of proteins at atomic-scale resolution. It is challenging to observe the dynamics of proteins using cryo-EM because of their large sizes and complex structural assemblies. A new deep-learning approach called DEFMap extracts the dynamics associated with the atomic fluctuations that are hidden in cryo-EM density maps.

It is a challenging task for any research field to screen the literature and determine what needs to be included in a systematic review in a transparent way. A new open source machine learning framework called ASReview, which employs active learning and offers a range of machine learning models, can check the literature efficiently and systemically.

In drug discovery and repurposing, systematic analysis of genome-wide gene expression of chemical perturbations on human cell lines is a useful approach, but is limited due to a relatively low experimental throughput. Computational, deep learning methods can help. In this work a graph neural network called Deep Chemical Expression is developed that can predict chemical-induced gene expression profiles. It is applied to identify drug repurposing candidates for COVID-19 treatments.

Self-driving vehicles must reliably detect the drivable area in front of them in any weather condition. An actively developed sensor approach is camera-based road segmentation, but it is limited by the visible spectrum. Radar-based approaches are a promising alternative and a new method extracts the drivable area from raw radar data by training a deep neural network using paired camera data, which can be labelled automatically using pretrained computer vision models.

Organic chemical reactions can be divided into classes that allow chemists to use the knowledge they have about optimal conditions for specific reactions in the context of other reactions of similar type. Schwaller et al. present here an efficient method based on transformer neural networks that learns a chemical space in which reactions of a similar class are grouped together.

Computational augmentation of microscopic images aims at reducing the need to chemically label or stain cells to extract information. The popular U-Net model often employed for these tasks uses mostly local information. A new method for augmenting microscopic images is presented that allows for global information to be used at each step of the process.