Volume 6 Issue 2, February 2024

Recent work has demonstrated important parallels between human visual representations and those found in deep neural networks. A new study comparing functional MRI data to deep neural network models highlights factors that may determine this similarity.

Training a machine learning model with multiple tasks can create more-useful representations and achieve better performance than training models for each task separately. In this Perspective, Allenspach et al. summarize and compare multi-task learning methods for computer-aided drug design.

Machine learning algorithms play important roles in medical imaging analysis but can be affected by biases in training data. Jones and colleagues discuss how causal reasoning can be used to better understand and tackle algorithmic bias in medical imaging analysis.

Recent years have seen many advances in deep learning models for protein design, usually involving a large amount of training data. Focusing on potential clinical impact, Garton et al. develop a variational autoencoder approach trained on sparse data of natural sequences of adenoviruses to generate large proteins that can be used as viral vectors in gene therapy.

Machine learning techniques are widely employed in chemical science, but are application specific and their development requires dedicated expertise. Jablonka and colleagues fine-tune the GPT-3 model and show that it can provide surprisingly accurate answers to a wide range of chemical questions.

Machine learning methods have made great advances in modelling protein sequences for a variety of downstream tasks. The representation used as input for these models has been primarily the sequence of amino acids. Outeiral and Deane show that using codon sequences instead can improve protein representations and lead to model performance.

Denoising low-counting statistics data in the presence of multiple, unknown noise profiles is a challenging task in scientific applications where high accuracy is required. Oppliger and colleagues train a deep convolutional neural network on pairs of experimental low- and high-noise X-ray diffraction data and demonstrate better performance on experimental noise filtering compared with the case of training on artificial data pairs.

Algorithmic decisions have a history of harming already marginalized populations. In an effort to combat these discriminative patterns, data-driven methods are used to comprehend these patterns, and recently also to identify disadvantaged communities to allocate resources. Huynh et al. analyse one of these tools and show a concerning sensitivity to input parameters that can lead to unintentional biases with substantial financial consequences.

Great advances in protein structure prediction have been made with recent deep learning-based methods, but proteins interact with their environment and can change shape drastically when binding to ligand molecules. To predict the 3D structure of these combined protein–ligand complexes, Qiao et al. developed a generative diffusion model with biophysical constraints and geometric deep learning.

Realistic quantum mechanical simulations are computationally costly to perform but can be approximated using neural network models. Li and colleagues propose a forward propagation method in lieu of traditional backpropagation to speed up these neural network-based approaches.

Deep learning language models have proved useful for both natural language and protein modelling. Similar to semantics in natural language, protein functions are complex and depend on the context of their environment, rather than on the similarity of sequences. Kulmanov and colleagues present an approach to frame function prediction as semantic entailment using a neuro-symbolic model to augment a large protein language model.

AI-enabled diagnostic applications in healthcare can be powerful, but study design is very important to avoid subtle issues of bias in the dataset and evaluation. Coppock et al. demonstrate how an AI-based classifier for diagnosing SARS-Cov-2 infection from audio recordings can seem to make predictions with high accuracy but shows much lower performance after taking into account confounders, providing insights in study design and replicability in AI-based audio analysis.

The “My Seizure Gauge” competition explored the challenge of forecasting epileptic seizures using non-invasive wearable devices without an electroencephalogram. The organizers and the winning team reflect on their experiences.