Research articles

The potential of deep learning in pathological prognosis has been hampered by limited interpretability in clinical applications. Liang and colleagues present a human-centric deep learning framework that supports the discovery of prognostic biomarkers in an interpretable way.

Generative models in cheminformatics depend on molecules being representable as structured data, such as the simplified molecular-input line-entry system (SMILES). Mokaya and colleagues investigated how the choice of representation influences the quality of generated compounds, and found that string-based representations can hinder performance in a curriculum learning setting.

Computational modelling of the interactions between T-cell receptors (TCRs) and epitopes is a crucial yet challenging scientific problem. Peng and colleagues develop a deep learning model to capture TCR–epitope binding patterns, providing useful insights for understanding TCR recognition.

Simulated data is an alternative to real data for medical applications where interventional data are needed to train AI-based systems. Gao and colleagues develop a model transfer paradigm to train deep networks on synthetic X-ray data and corresponding labels generated using simulation techniques from CT scans. The approach establishes synthetic data as a viable resource for developing machine learning models that apply to real clinical data.

Stochastic reaction networks involve solving a system of ordinary differential equations, which becomes challenging as the number of reactive species grows, but a new approach based on evolving a variational autoregressive neural network provides an efficient way to track time evolution of the joint probability distribution for general reaction networks.

Computational models can help predict metabolic profiles of microbial communities such as human gut microbiomes or environmental microbiomes, but they lack generalizability and interpretability. To address this challenge, Wang et al. report a deep learning approach for metabolic profile prediction called mNODE that incorporates a neural network module with hidden layers described by ordinary differential equations.

Various post-hoc interpretability methods exist to evaluate the results of machine learning classification and prediction tasks. To better understand the performance and reliability of such methods, which is particularly necessary in high-risk applications, Turbe et al. have developed a framework for quantitative comparison of post-hoc interpretability approaches in time-series classification.

Metal–organic frameworks are of high interest for a range of energy and environmental applications due to their stable gas storage properties. A new machine learning approach based on a pre-trained multi-modal transformer can be fine-tuned with small datasets to predict structure-property relationships and design new metal-organic frameworks for a range of specific tasks.

Neuro-symbolic artificial intelligence approaches display both perception and reasoning capabilities, but inherit the limitations of their individual deep learning and symbolic artificial intelligence components. By combining neural networks and vector-symbolic architectures, Hersche and colleagues propose a neuro-vector-symbolic framework that can solve Raven’s progressive matrices tests faster and more accurately than other state-of-the-art methods.

Explanatory interactive machine learning methods have been developed to facilitate the learning process between the machine and the user. Friedrich et al. provide a unification of various explanatory interactive machine learning methods into a single typology, and present benchmarks for evaluating such methods.

Machine learning methods can predict and recognize binding patterns between T-cell receptors and human antigens, but they struggle with antigens for which no or little data exist regarding interactions with the immune system. A new method called PanPep based on meta-learning can learn quickly on new binding prediction tasks and accurately predicts pairing between T-cell receptors and new antigens.

High-quality annotation of datasets is critical for machine-learning-based biomedical image analysis. However, a detailed examination of recent image competitions reveals a gap between annotators’ needs and quality of labelling instructions. It is also found that annotator performance can be substantially improved by providing exemplary images.

Training a deep neural network can be costly but training time is reduced when a pre-trained network can be adapted to different use cases. Ideally, only a small number of parameters needs to be changed in this process of fine-tuning, which can then be more easily distributed. In this Analysis, different methods of fine-tuning with only a small number of parameters are compared on a large set of natural language processing tasks.

Reinforcement learning is a powerful technique to learn complex behaviours, but in the context of self-driving vehicles it might result in unsafe behaviour in previously unseen situations. Cao et al. create a confidence-aware method that improves through reinforcement learning but reverts to safe behaviour when a situation is new.

Developing proprioception systems for flexible structures such as soft robots is a challenge. Hu et al. report a stretchable e-skin for soft robot proprioception. Combined with deep learning, the e-skin enables high-resolution 3D geometry reconstruction of the soft robot and can be applied in many scenarios, such as human–robot interaction.

The mechanical signals of the laryngeal vocal organ have not been well utilized by human speech processing technology. The authors develop a prototype of a wearable artificial throat that can sense speech- and vocalization-related actions. The results suggest a new technological pathway for speech recognition and interaction systems.

When learning a causal model from data, deriving counterfactual examples from the model can help to evaluate how plausible the mechanisms are and create hypotheses that can be tested with new data. Vlontzos and colleagues develop a deep learning-based method for answering counterfactual queries that can deal with categorical variables, rather than only binary ones, using the notion of ‘counterfactual ordering’.

Co-designing hardware platforms and neural network software can help improve the computational efficiency and training affordability of deep learning implementations. A new approach designed for graph learning with echo state neural networks makes use of in-memory computing with resistive memory and shows up to a 35 times improvement in the energy efficiency and 99% reduction in training cost for graph classification on large datasets.

Disease phenotypes can be predicted from genetic profiles, but diseases with complex, non-additive interactions between genes are hard to disentangle. An approach called DiseaseCapsule makes use of capsule networks to identify the hierarchical structure in genomic data and can predict complex diseases such as amyotrophic lateral sclerosis with high accuracy.

Predicting drug–target interaction with computational models has attracted a lot of attention, but it is a difficult problem to generalize across domains to out-of-distribution data. Bai et al. present here a method that aims to model local interactions of proteins and drug molecules while being interpretable and provide cross-domain generalization.