The directed evolution of biomolecules is an iterative process. Although advancements in language models have expedited protein evolution, effectively evolving RNA remains a challenge. RNA aptamers, selected for their binding properties, provide an ideal system to address this challenge, yet traditional aptamer discovery still relies on labor-intensive, multi-round screening. Here we introduce GRAPE-LM (generator of RNA aptamers powered by activity-guided evolution and language model), a generative artificial intelligence framework designed for the one-round evolution of RNA aptamers. GRAPE-LM integrates a transformer-based conditional autoencoder with nucleic acid language models and is guided by CRISPR−Cas-based aptamer screening data derived from intracellular environments. We validate GRAPE-LM on three disparate targets: the human T cell receptor CD3ε, the receptor-binding domain of the SARS-CoV-2 spike protein and the human oncogenic transcription factor c-Myc (an intracellular disordered protein). GRAPE-LM, informed with only a single round of CRISPR−Cas-based screening, successfully obtains RNA aptamers that outperform those driven from multiple rounds of human selection and optimization. Combining generative AI and one round of wet lab evolution generates high-affinity RNA aptamers.

The extent to which microbial processes control soil organic carbon (SOC) dynamics remains uncertain. Carbon use efficiency (CUE), that is, the fraction of assimilated carbon allocated to growth, has been used as a key parameter but its relationship with SOC reflects carbon partitioning rather than the absolute magnitude of microbial fluxes. The microbial growth rate could provide a more mechanistic link to SOC accumulation because it quantifies biomass production and reflects necromass formation. Here we combine a global ¹⁸O–H2O dataset (n = 268 paired observations) with outputs from four land surface models to test whether growth rate predicts SOC more strongly than CUE. In the incubation experiments, growth rates are more closely associated with SOC than CUE, although soil properties and climate explain equal or greater variance. Models reproduce the stronger role of growth rate over CUE but tend to underestimate the abiotic controls. The models also emphasize CUE as the main predictor of the SOC-to-net primary production ratio, in contrast to observations, which indicates the soil’s capacity to retain plant carbon inputs. Together, these findings identify the microbial growth rate as a diagnostic that can help bridge models with empirical data and guide a more balanced representation of microbial and mineral controls in SOC projections. Microbial carbon use efficiency is a strong predictor of soil organic carbon stocks. Here the authors reveal that the microbial growth rate is a more reliable and informative predictor, and that modelling approaches tend to overemphasize the role of biotic over abiotic controls compared to empirical data.

Owing to their natural origin and biocompatibility, extracellular vesicles (EVs) are being recognized as next-generation vehicles for targeted drug delivery. Despite their potential as therapeutic carriers, EVs suffer from heterogeneity, low yields, limited cargo loading efficiency and rapid clearance by the mononuclear phagocyte system. Since the first EV-based clinical trial in 2005, more than 100 clinical trials have investigated the use of EVs as therapeutics and drug carriers. Despite this, no EV-based therapies have received regulatory approval to date. This gap between preclinical research activity and clinical translation underscores persistent scientific challenges and regulatory hurdles that continue to impede the advancement of EV-based therapeutics. In this Review, we examine the research articles published in the field between 2012 and 2024 (38,177 articles), highlighting key developments, persistent challenges and evolving assumptions. We review the current EV landscape and clinical trials, focusing on their organotropism and use as carriers for therapeutics. We compare their advantages and limitations in relation to other nanoparticles, such as lipid nanoparticles and liposomes, and examine how labelling strategies and cell sources influence EV biodistribution. Finally, we outline translational considerations for EV-based therapeutics and propose additional reporting standards, complementing the MISEV 2023 guidelines. Extracellular vesicles (EVs) are emerging as biocompatible carriers for targeted drug delivery, yet challenges in yield, cargo loading and biodistribution persist. In this Review, key trends from over a decade of research are analysed, comparing EVs with lipid-based systems and outlining strategies for improving therapeutic translation and reporting standards.

Demand for recombinant proteins is rapidly growing, driven by their use as biotherapeutics, vaccine components, industrial enzymes, and food ingredients. The growing market requires novel strategies for increasing protein production in cellular hosts. Systems-level frameworks have been used to improve production, but have had difficulty relating complex cellular pathways with protein expression. Here, we demonstrate a method for mapping relationships between gene expression signatures and carbon source-related phenotypes related to recombinant protein production.  Our approach induces systematic perturbations in cultures of K. phaffii using varied co-feeds of carbon sources. The different carbon sources significantly impacted cell growth, specific productivity, and transcriptional states. With these data, we identified metagenes for both immunoglobulin G1 monoclonal antibody (IgG1) and Variable domain on a heavy chain (VHH) antibody that explained significant transcriptomic variance. These metagenes strongly associated with two phenotypes: production of recombinant protein-to-biomass ratio, and response to methanol induction. We used these results to identify and knockout 31 novel gene targets for which expression inversely correlated with productivity. Nine of these genes improved productivity of IgG1 by up to 3x and ten genes increased productivity of VHH by up to 1.7x. Many of these genes are involved in the modulation and progression of the cell cycle but interestingly, disruption had little to no impact on cell growth.  This study establishes a framework for relating gene signatures to complex cellular phenotypes, providing a robust methodology for assessing production processes and identifying new targets for cellular engineering. While the identified specific metagenes depend on the complexity and structure of the recombinant protein produced, this framework is extensible across diverse proteins and potentially other host organisms. These signatures may serve as scale-independent, cellular-level metrics for traits like efficiency of production of recombinant proteins, facilitating the translation of findings across different scales and cultivation modes. Furthermore, this framework enables the identification of novel targets for genomic modifications that can improve strain performance, offering a predictive tool for the rational design of high-performing microbial cell factories. 

Genetic interaction and protein-protein interaction networks have proved among the most important tools for inference of gene function and genotype-phenotype relationship. It remains unclear, however, how these two networks are related to each other. Here we demonstrate that the strengths of epistatic genetic interactions between two genes strongly correlate with the binding free energies of interactions between their cognate proteins, for both yeast and human genomes. Consequently, we show that genetic interaction and protein-protein interaction networks can reciprocally predict each other. Further, we observe that functional divergence and redundancy in duplicated genes (paralogs) are reflected in the binding affinity of their interactions with other proteins and in their genetic interaction strengths. Finally, we demonstrate that the overall topologies of genetic interaction and protein-protein interaction networks significantly overlap in two topological features: modules, in which genes/proteins are organized by common function, and connectors, which link modules to each other. This opens avenues for integrated network approaches in understanding cellular processes and flow of genomic information. Genetic and protein-protein interaction networks can be used to infer gene function and genotype-phenotype relationships. Here, the authors show that genetic interaction patterns reflect protein binding affinities.

Treatment of sensitive bacteria with beta‐lactam antibiotics often leads to two salient population‐level features: a transient increase in total population biomass before a subsequent decline, and a linear correlation between growth and killing rates. However, it remains unclear how these population‐level responses emerge from collective single‐cell responses. During beta‐lactam treatment, it is well‐recognized that individual cells often exhibit varying degrees of filamentation before lysis. We show that the cumulative probability of cell lysis increases sigmoidally with the extent of filamentation and that this dependence is characterized by unique parameters that are specific to bacterial strain, antibiotic dose, and growth condition. Modeling demonstrates how the single‐cell lysis probabilities can give rise to population‐level biomass dynamics, which were experimentally validated. This mapping provides insights into how the population biomass time‐kill curve emerges from single cells and allows the representation of both single‐ and population‐level responses with universal parameters.

Intrinsically disordered proteins are ubiquitous in biological systems and play essential roles in a wide range of biological processes and diseases. Despite recent advances in high-resolution structural biology techniques and breakthroughs in deep learning-based protein structure prediction, accurately determining structural ensembles of IDPs at atomic resolution remains a major challenge. Here, we introduce bAIes, a Bayesian framework that integrates AlphaFold2 predictions with physico-chemical molecular mechanics force fields to generate accurate atomic-resolution ensembles of IDPs. We show that bAIes produces structural ensembles that match a wide range of high- and low-resolution experimental data across diverse systems, achieving accuracy comparable to atomistic molecular dynamics simulations but at a fraction of their computational cost. Furthermore, bAIes outperforms state-of-the-art IDP models based on coarse-grained potentials as well as deep-learning approaches. Our findings pave the way for integrating structural information from modern deep-learning approaches with molecular simulations, advancing ensemble-based understanding of disordered proteins. Here, the authors present bAIes, which integrates AlphaFold2 information into a molecular dynamics framework to efficiently generate atomic-resolution ensembles of intrinsically disordered proteins.