Articles in 2022

Large language models have recently emerged with extraordinary capabilities, and these methods can be applied to model other kinds of sequence, such as string representations of molecules. Ross and colleagues have created a transformer-based model, trained on a large dataset of molecules, which provides good results on property prediction tasks.

Liquid chromatography–tandem mass spectrometry (LC-MS²) provides high-throughput screening of molecules with a large number of features. But these features are difficult to associate with specific molecular structures of each sample. To improve structure prediction from these features, Bach et al. propose a machine learning model trained to also take into account stereochemistry to combine the different kinds of features provided by LC-MS².

There is growing interest in using sophisticated machine learning models for the prediction of molecular properties, such as potency of novel drugs. However, Janela and Bajorath show that simple nearest-neighbour analysis meets or exceeds the accuracy of state-of-the-art complex machine learning methods and that randomized prediction models still reproduce compound potency values within an order of magnitude.

The design of legged robots with agility and speed is challenging. The authors present a method with reinforcement learning-based controllers for locomotion control of quadruped robots. The pipeline achieves improvements in performance, such as running speed.

Predicting RNA degradation is a fundamental task in designing RNA-based therapeutics. Two crowdsourcing platforms, Kaggle and Eterna, united to develop accurate deep learning models for RNA degradation on a timescale of 6 months.

The problem of reconstructing full-field quantities from incomplete observations arises in various real-world applications. Güemes and colleagues propose a super-resolution algorithm based on a generative adversarial network that can achieve reconstruction of the underlying field from random sparse measurements without requiring full-field high-resolution training data.

Evolutionary computation is a very active field of research, with an ever-growing number of metaheuristic optimization algorithms being published. A serious problem plaguing the field is the use of inadequate benchmarks. Kudela exposes the issue and provides recommendations that can help to fairly evaluate and compare new methods.

Learning minimal representations of dynamical systems is essential for mathematical modelling and prediction in science and engineering. Floryan and Graham propose a deep learning framework able to estimate accurate global dynamical models by sewing together multiple local representations learnt from high-dimensional time-series data.

Advances in ultra-widefield retinal imaging have created a need for automated disease detection. Engelmann and colleagues develop a deep learning model for the detection of retinal diseases. They evaluate it under more realistic conditions than has been considered previously and investigate what regions of ultra-widefield images are important for the performance of such a model.

In recent years, deep learning techniques have enhanced the possibility to extract useful, high-resolution physical information from electron and scanning probe microscopy images. AtomAI, an open-source software package, can accelerate this process by bringing deep learning and simulation tools into a single framework for a range of instruments.

Mutation rates are crucial for genetic and evolutionary analyses. Fang et al. present a generalizable deep learning method to build fine-scale mutation rate maps with DNA sequences as input, which can benefit analyses reliant on mutation rates.

A challenge for any machine learning system is to continually adapt to new data. While methods to address this issue are developed, their performance is hard to compare. A new framework to facilitate benchmarking divides approaches into three categories, defined by whether models need to adapt to new tasks, domains or classes.

Generative AI that transforms candidate molecules into potent drugs could pave the way towards better, faster drug design. Chan and co-workers developed a self-contrastive learning framework that helps AI models overcome data biases and provide explainable insights for augmented interactive design.

The lack of generalizability and reproducibility of machine learning models in medical applications is increasingly recognized as a substantial barrier to implementing such approaches in real-world clinical settings. Highlighting this issue, Jie Cao et al. aim to reproduce a recent acute kidney injury prediction model, and find persistent discrepancies in model performance in different subgroups.

A promising area for deep learning is in modelling complex physical processes described by partial differential equations (PDEs), which is computationally expensive for conventional approaches. An operator learning approach called DeepONet was recently introduced to tackle PDE-related problems, and in new work, this approach is extended with transfer learning, which transfers knowledge obtained from learning to perform one task to a related but different task.

Predicting the properties of a molecule from its structure with high accuracy is a crucial problem in digital drug design. Instead of sequence features, Zeng and colleagues use an image representation of a large collection of bioactive molecules to pretrain a model that can be fine-tuned on specific property prediction tasks.

Predicting disease progression is an important medical problem, but it can be challenging for end-to-end machine learning approaches. Han and colleagues demonstrate that generative models can work together with medical experts to jointly predict the progression of a disease, osteoarthritis.

Physical dynamical processes can be modelled with differential equations that may be solved with numerical approaches, but this is computationally costly as the processes grow in complexity. In a new approach, dynamical processes are modelled with closed-form continuous-depth artificial neural networks. Improved efficiency in training and inference is demonstrated on various sequence modelling tasks including human action recognition and steering in autonomous driving.

Computational methods are important for interpreting missense variants in genetic studies and clinical testing. Zhang and colleagues develop a method based on graph attention neural networks to predict pathogenic missense variants. The method pools information from functionally correlated positions and can improve the interpretation of missense variants.

To enable electrification of cities, thereby decarbonizing energy grids, it is essential to predict electricity consumption with high spatio-temporal accuracy. To reduce the need for large amounts of training data, an active deep learning approach is developed to make accurate forecasts of electric load profiles at the scale of single buildings.