Insertions and deletions (indels) are difficult to study because they change protein sequence length, limiting existing tools. Here, IndeLLM is presented: a framework that leverages protein language models to assess indel effects in a more interpretable and generalizable way. The method maps indel impact to protein regions, improves predictive accuracy with a Siamese network, and provides guidelines for transfer learning. IndeLLM offers new opportunities for indel annotation and disease-related variant analysis.

When AI surpasses human performance, what can humans offer? We demonstrate that the performance of teams increases by integrating human judgments with those of machines. Integration is achieved by a straightforward regression approach that combines team members' confidence-weighted judgments.

This study gives an example of how complex computations in neural networks can emerge from simple physical principles. Training a model to optimize internal energy efficiency while performing eye movements suffices for predictive remapping to emerge. The model learns to translate eye movements into an allocentric reference frame. Based on this reference frame, it learns to predict and inhibit the next fixation.

Rueda et al. argue that the concept of dignity is problematic for AI ethics due to its complexity, ambiguity, and biased usage. While agreeing on many points, we propose that adding the necessary prec...

Antibodies are important for human health, but their diversity makes computational prediction of their binding properties challenging. Through contrastive fine-tuning of antibody language models on millions of antibody pairs, the authors enable accurate prediction of epitope overlap even among sequence-diverse antibodies, providing powerful new tools for therapeutic antibody discovery.

The authors analyze how 14 environmental research infrastructures (RIs) implement the FAIR (Findable, Accessible, Interoperable, and Reusable) principles in Italy. The study uncovers diverse practices across the atmosphere, marine, biosphere, and geosphere subdomains. Results reveal not only strong heterogeneity but also emerging convergence, particularly in the marine subdomain. By highlighting transferable FAIR strategies, this work can guide other scientific communities interested in FAIR compliance and support the development of new digital ecosystems within and beyond Italy.

This perspective surveys three-factor learning, a key mechanism for brain-inspired spiking neural networks. Adding a third, global signal, analogous to neuromodulators, allows these AI systems to learn and adapt more effectively. This overview connects recent advances in neuroscience and machine learning, outlining a path toward more powerful AI.

We live in an era of escalating sustainability challenges, such as exceeding the global 1.5°C threshold in 2024, the collapse of talks to address the widespread plastic crisis, and increasing wildfire...

As artificial intelligence matures, the impact it might have on how society functions is being actively pondered. In this opinion, through uniform-binomial mixtures, the author sheds quantitative ligh...

Data constitute machine learning’s fuel. Unfortunately, real-world data are often imbalanced. This poses a huge challenge to machine learning, as rare cases are hard to learn. So far, no adequate measure has been available to assess the impact of imbalance on machine learning’s performance. As a remedy for this omission, the authors introduce the status of imbalance (SIMBA), a reliable, robust, and generic imbalance measure that outperforms existing ones.

Brain-computer interfaces (BCIs) offer a promising avenue to restore motor function by converting neural activity into control signals for external devices. While current BCI technologies have successfully demonstrated brain-directed control of robotic arms, they still face significant limitations in dexterous hand movement control. This study proposes a novel approach called SynergyNet to decode complex hand movements directly from brain signals, facilitating high-performance BCI control with fine movements.

In this review, the authors explore the transformative impact of large language models (LLMs) on drug discovery and development. They detail how LLMs can potentially accelerate our understanding of…

Quaternions, a type of hypercomplex number, are particularly useful in handling three-dimensional data, i.e., color images. Within the feature-rich hypercomplex setting, image-processing workflows can be realized for natural and biomedical images enabling alternative visual representations, offering effective solutions to current problems in computer vision and digital pathology, and generally expanding the scope and impact of hypercomplex image processing across a wide range of applications.

Transformer models are adaptable to off-the-beaten-path molecular design tasks, such as protein-sequence-based compound design. The authors use these predictions as a model system to unravel the learning characteristics of transformers. The analysis reveals that the transformer predictions are statistically driven and that the models do not learn protein-ligand interactions or other biologically relevant information. Instead, compound memorization plays an important role. These findings caution against over-interpretation of sequence-based generative compound design using transformer models.

This study explores AI-generated writing adoption across consumer complaints, corporate press releases, job postings, and international organization communications, revealing widespread use following ...

This study explores AI-generated writing adoption across consumer complaints, corporate press releases, job postings, and international organization communications, revealing widespread use following ChatGPT’s release. Findings highlight significant generative AI integration by small firms, corporations, and global institutions, emphasizing its transformative impact on communication practices, business strategies, and public policy.

Visual quality assessment is crucial for both human perception and artificial vision systems. This study reveals how the human brain naturally perceives image quality through distinct neural pathways. Functional magnetic resonance imaging results show that while low-level visual areas are highly sensitive to distortions, higher-level regions compensate to maintain recognition, much like how AI systems might “fill in” missing details. These findings not only deepen our understanding of human vision but also provide biological inspiration for developing more robust image-processing algorithms.

The vast microbial “dark proteome” remains largely uncharacterized. Nelson et al. present LA4SR, a transformer framework that classifies algal sequences 10,000× faster than traditional methods with near-perfect recall. Trained on 77 million sequences including synthetic chimeras, the system can robustly classify unknown sequences with natural or scrambled start and stop information. They also developed an accompanying suite of explainability software to reveal biologically meaningful patterns linking amino acid patterns to evolutionary affiliations.

This study introduces a language-model-driven agent system that captures and applies laboratory knowledge to autonomously plan, execute, and analyze experiments. When the agent system was demonstrated on quantum computing tasks, the agents achieved performance on par with expert researchers. By enabling intelligent coordination and decision-making in complex lab settings, this approach paves the way for more flexible, scalable automation in experimental science.