Articles in 2024

Identifying compounds in tandem mass spectrometry requires extensive databases of known compounds or computational methods to simulate spectra for samples not found in databases. Simulating tandem mass spectra is still challenging, and long-range connections in particular are difficult to model for graph neural networks. Young and colleagues use a graph transformer model to learn patterns of long-distance relations between atoms and molecules.

The 5′ untranslated region is a critical regulatory region of mRNA, influencing gene expression regulation and translation. Chu, Yu and colleagues develop a language model for analysing untranslated regions of mRNA. The model, pretrained on data from diverse species, enhances the prediction of mRNA translation activities and has implications for new vaccine design.

Using machine learning methods to model interatomic potentials enables molecular dynamics simulations with ab initio level accuracy at a relatively low computational cost, but requires a large number of labelled training data obtained through expensive ab initio computations. Cui and colleagues propose a geometric learning framework that leverages self-supervised learning pretraining to enhance existing machine learning based interatomic potential models at a negligible additional computational cost.

The area under the receiver operating characteristic curve (AUROC) of the test set is used throughout machine learning (ML) for assessing a model’s performance. However, when concordance is not the only ambition, this gives only a partial insight into performance, masking distribution shifts of model outputs and model instability.

In early 2023, Bai and colleagues presented DrugBAN, an interpretable method for drug–target prediction. In this Reusability Report, Xu and colleagues reproduce the original findings and provide a careful exploration of cross-domain adaptability.

Generative models for chemical structures are often trained to create output in the common SMILES notation. Michael Skinnider shows that training models with the goal of avoiding the generation of incorrect SMILES strings is detrimental to learning other chemical properties and that allowing models to generate incorrect molecules, which can be easily removed post hoc, leads to better performing models.

After several decades of developments in AI, has the inspiration that can be drawn from neuroscience been exhausted? Recent initiatives make the case for taking a fresh look at the intersection between the two fields.

AI methods can discover new antibiotics but existing methods have limitations. Swanson et al. develop a generative AI model that learns to design molecules that are easy to synthesize. The authors apply the model to design and validate novel antibiotics against the bacterial pathogen Acinetobacter baumannii.

An emerging research area in AI is developing multi-agent capabilities with collections of interacting AI systems. Andrea Soltoggio and colleagues develop a vision for combining such approaches with current edge computing technology and lifelong learning advances. The envisioned network of AI agents could quickly learn new tasks in open-ended applications, with individual AI agents independently learning and contributing to and benefiting from collective knowledge.

As the impacts of AI on everyday life increase, guidelines are needed to ensure ethical deployment and use of this technology. This is even more pressing for technology that interacts with groups that need special protection, such as children. In this Perspective Wang et al. survey the existing AI ethics guidelines with a focus on children’s issues, and provide suggestions for further development.

Although federated learning is often seen as a promising solution to allow AI innovation while addressing privacy concerns, we argue that this technology does not fix all underlying data ethics concerns. Benefiting from federated learning in digital health requires acknowledgement of its limitations.

Foundation models have transformed artificial intelligence by training on vast amounts of broad unlabelled data. Pai et al. present a foundation model leading to more accurate, efficient and robust cancer imaging biomarkers, especially in use cases with small training datasets.

AI tools such as ChatGPT can provide responses to queries on any topic, but can such large language models accurately ‘write’ molecules as output to our specification? Results now show that models trained on general text can be tweaked with small amounts of chemical data to predict molecular properties, or to design molecules based on a target feature.

Deep learning generative approaches have been used in recent years to discover new molecules with drug-like properties. To improve the performance of such approaches, Yang et al. add chemical binding knowledge to a deep generative framework and demonstrate, including by wet-lab verification, that the method can find valid molecules that successfully bind to target proteins.

Genome-wide association studies allow connecting genomic information with complex traits. Rodrigo Bonazzola et al. develop a framework consisting of several deep learning tools to improve the discoverability of genes that influence specific geometric features of the heart.

Can non-state multinational tech companies counteract the potential democratic deficit in the emerging global governance of AI? We argue that although they may strengthen core values of democracy such as accountability and transparency, they currently lack the right kind of authority to democratize global AI governance.

Visual representations are thought to develop from visual experience and inductive biases. Orhan and Lake show that modern machine learning algorithms can learn visual knowledge from a few hundred hours of longitudinal headcam recordings collected from young children during the course of early development, without strong inductive biases.

This Reusability Report examines a recently published deep learning method PENCIL by Ren et al. for identifying phenotype populations in single-cell data. Cao et al. reproduce here the main results, analyse the sensitivity of the method to model parameters and describe how the method can be used to create a signature for immunotherapy response markers.