Vaccines are the most effective tool in preventing and managing infectious diseases. One of the critical challenges in vaccine development is the selection of suitable target antigens from the thousands of proteins produced by pathogens. Artificial intelligence is anticipated to play a significant role in addressing this challenge. In this study, we develop a framework termed PLGDL for protective antigen prediction that employs Protein Language and Geometric Deep Learning models. This framework leverages both primary sequence features and three-dimensional structural features of protein antigens, thereby reducing the biases associated with manually curated features. Our integrated model exhibits robustness across both constructed and public datasets and is applicable to viruses, bacteria, and eukaryotic pathogens. Notably, when applied to the ongoing Mpox outbreak, our model not only quickly identifies multiple known antigens but also discovers a protective antigen: G10R. Here, our study provides a high-performance screening tool for protective vaccine antigen prediction by synergistically utilizing the capabilities of protein language and geometric deep learning models, providing substantive insights and methodological advancements for rapid vaccine development. Vaccines are the most effective tool in managing infectious disease and characterizing features of protective epitopes could help in prediction methods. Here the authors use protein language and geometric deep learning frameworks to investigate primary sequence features and structural features to identify and predict potential antigens, showing prediction of a protective mpox epitope using this method.

Heterogeneity within clonal cell populations remains a critical bottleneck within bioprocess engineering, notably by undermining bioproduction yields. Efforts to mitigate its impact have, however, been hampered by technological difficulties quantifying metabolism at the single-cell level. Here, we propose a framework based on single-cell biosensor analysis that enables robust characterization of cell’s metabolic states, leveraging it to detect and isolate isogeneic heterogeneity in response to environmental perturbations and within microbial cell factories. We identify acute and gradual glucose depletion to induce differentiation of metabolically distinct subpopulations and reveal these subpopulations to exhibit differential production capabilities, with lower intracellular pH subpopulations exhibiting enhanced product accumulation within violacein-producing strains but reduced yields within lycopene-producing strains. Lastly, we highlight galactose cultivation as a method to modulate subpopulation dynamics towards higher-producing lycopene phenotypes. Altogether, our research provides insights into subpopulation differentiation and establishes promising avenues for the engineering of more robust and higher-producing strains. Heterogeneity within clonal cell populations affects bioprocess engineering. Here, the authors report a biosensor-based toolkit to investigate phenotypic heterogeneity in engineered yeast, reveal pH-based subpopulations and metabolite production states, and modulate/shift subpopulation dynamics to increase lycopene production.

Richard Lenski traces the legacy of Escherichia coli and how science is evolving to use this model organism in new ways.

Point-of-use diagnostics based on allosteric transcription factors (aTFs) are promising tools for environmental monitoring and human health. However, biosensors relying on natural aTFs rarely exhibit the sensitivity and selectivity needed for real-world applications, and traditional directed evolution struggles to optimize multiple biosensor properties at once. To overcome these challenges, we develop a multi-objective, machine learning (ML)-guided cell-free gene expression workflow for engineering aTF-based biosensors. Our approach rapidly generates high-quality sequence-to-function data, which we transform into an augmented paired dataset to train an ML model using directional labels that capture how aTF mutations alter performance. We apply our workflow to engineer the aTF PbrR as a point-of-use diagnostic for lead contamination in water. We tune the sensitivity of PbrR to sense at the U.S. Environmental Protection Agency (EPA) action level for lead and modify the selectivity away from zinc, a common metal found in water supplies. Finally, we show that the engineered PbrR functions in freeze-dried cell-free reactions, enabling a diagnostic capable of detecting lead in drinking water down to ~5.7 ppb. Our ML-driven, multi-objective framework powered by directional tokens can generalize to other biosensors and proteins, accelerating the development of synthetic biology tools for biotechnology applications. Allosteric transcription factors (aTFs) are promising tools for environmental and human health monitoring. Here the authors develop a multi-objective, machine learning-guided method to engineer an aTF-based portable diagnostic for environment sensing of lead in drinking water at the legal limit.

The production of recombinant proteins in E. coli is often hampered by the formation of inclusion bodies. While fusion tags can enhance solubility, existing systems are hampered by a lack of standardization, with tags scattered across disparate plasmid backbones and inconsistent cloning sites, complicating parallel screening. To address this, we constructed a standardized series of expression vectors, termed pNX, by incorporating nine small fusion tags (SUMO, LD, ACP, BCCP, GB1, Fh8, SmbP, TolA, and TrxA) into a uniform pET-28b backbone. Each pNX vector features an identical configuration: a T7 promoter, an N-terminal fusion tag, a synthetic linker, a TEV protease cleavage site, and a multiple cloning site (MCS) flanked by dual 6×His tags. We evaluated this system using four model proteins (EcFabG, eGFP, XccXanA2, and XccXanL). Our results showed that specific tags significantly improved both the expression level and solubility of the target proteins without compromising their biological activity. Notably, the lipoyl domain (LD) was identified, to our knowledge for the first time, as an effective solubility enhancer. The standardized MCS enabled rapid, parallel cloning, facilitating the efficient screening of optimal fusion partners. The pNX vector series provides a versatile and efficient platform for enhancing the soluble expression of challenging recombinant proteins in E. coli, streamlining the empirical identification of ideal fusion tags.

The discovery of the antiviral-responsive RNA editing system is a significant advance in eukaryotic biology. It reveals new dimensions of RNA editing diversity and fungal antiviral strategies, and provides compelling evidence for the adaptive significance of RNA editing in host-virus coevolution. This discovery establishes a novel model for investigating both the molecular mechanisms of RNA editing and the evolutionary dynamics of altruistic defense.

Biologically meaningful interpretation of transcriptomic datasets remains challenging, particularly when context-specific gene sets are either unavailable or too generic to capture the underlying biology. We here present InCURA, an integrative clustering strategy based on transcription factor (TF) motif occurrence patterns in gene promoters. InCURA takes as input lists of (i) all expressed genes, used solely to identify dataset-specific expressed TFs, and (ii) differentially regulated genes (DRGs) used for clustering. Promoter sequences of DRGs are scanned for TF binding motifs, and the resulting counts are compiled into a gene-by-TFBS matrix. InCURA then uses unsupervised clustering to infer gene modules with shared predicted regulatory input. Applying InCURA to diverse biological datasets, we uncovered functionally coherent gene modules revealing upstream regulators and regulatory programs that standard enrichment or co-expression analyses fail to detect. In summary, InCURA provides a user-friendly, regulation-centric tool for dissecting transcriptional responses, particularly in settings lacking context-specific gene sets.

Cell-to-cell variability often limits the efficiency of microbial bioproduction, yet how individual cells fluctuate over time and how these fluctuations shape population-level output remain unclear. To address this issue, we tracked a heterologous betaxanthin pathway in Escherichia coli using microfluidics-assisted time-lapse microscopy, allowing simultaneous measurement of fluctuations in betaxanthin, its biosynthetic enzyme DOD and growth across generations. Here we show that over 50% of high betaxanthin producers become medium or low producers after two divisions. Betaxanthin variation primarily originates from DOD noise, with a smaller contribution from growth rate fluctuations. We further develop a stochastic model to explore various control circuits and find that pathway enzyme or metabolite-based growth selection strategies are most effective in enhancing production. We experimentally validate the model by coupling enzyme expression to nutrient availability, which enriches high producers and boosts titer by 4.4-fold. Our results highlight key sources of metabolic heterogeneity and provide a framework for designing robust microbial processes. Cell-to-cell variability limits efficient microbial production. Here, the authors track single cells to reveal enzyme noise as the main source of bioproduction variation, and by coupling growth to pathway performance, they selectively enrich high producers and substantially boost overall titres.

Light-sensitive proteins allow organisms to perceive and respond to their environment, and have diversified over billions of years. Among these, Light-Oxygen-Voltage (LOV) domains are widespread photosensors that control diverse physiological processes and are increasingly used in optogenetics. Yet, the evolutionary constraints that shaped their protein dynamics and thereby their functional diversity remain poorly resolved. Here we systematically characterize the dynamics of 21 natural LOV core domains, significantly extending the spectroscopically resolved catalog through the addition of 18 previously unstudied variants. Using time-resolved spectroscopy, we uncover an exceptional kinetic diversity spanning from picoseconds to days and identify distinct functional clusters within the LOV family. These clusters reflect evolutionary branching, including a divergence of  1.0 billion years between investigated LOV variants from plants and 0.4 billion years of separation within one of these functional clusters. Individual variants with extreme photocycles emerge as promising anchor points for optogenetic applications, ranging from highly efficient adduct formation to ultrafast recovery. Beyond natural diversity, we introduce a LOV domain generated by artificial intelligence-guided protein design. Despite being sequentially remote from its maternal template, this variant retains core photocycle function while exhibiting unique biophysical properties, thereby occupying a new region on the biophysical landscape. Our work emphasizes how billions of years of evolution defined LOV protein dynamics, and how protein design can expand this repertoire, engineering next-generation optogenetic tools.

Although virus ecogenomics has expanded access to and understanding of the virosphere, existing classification tools lack taxonomic resolution and are unable to scale to modern discovery-based datasets or classify previously unknown sequence space. Here we develop vConTACT3—a machine learning-based tool that improves scalability and accuracy of virus taxonomy. By optimizing gene-sharing thresholds and leveraging adaptive, realm-specific cut-offs, vConTACT3 expands classification to both eukaryote and prokaryote viruses for four of the six officially recognized realms, and establishes accurate hierarchical taxonomy from genus to order. Specifically, vConTACT3 achieves >95% agreement with official taxonomy for 35,545 and 13,524 public prokaryotic and eukaryotic virus genomes, respectively, to surpass vConTACT2 across most realms, while still uniquely classifying previously uncharacterized taxa, and doing so even faster. vConTACT3 application provides taxonomy assignments for tens of thousands of unclassified taxa rapidly, automatically and systematically; evaluates virus sequence space to reveal support for fewer taxonomic ranks than currently available and identifies taxonomically challenging areas across the virosphere. vConTACT3 enables multirank, large-scale classification of eukaryotic and prokaryotic viruses.

Predicting changes in protein thermostability caused by amino acid substitutions is essential for understanding human diseases and engineering proteins for practical applications. While recent protein generative models demonstrate impressive zero-shot performance in predicting various protein properties without task-specific training, their strong unsupervised prediction ability remains underexploited to improve protein stability prediction. We present SPURS, a deep learning framework that rewires and integrates two complementary protein generative models–a protein language model and an inverse folding model–and reprograms this unified framework for stability prediction through supervised fine-tuning on mega-scale thermostability data. SPURS delivers accurate, efficient, and scalable stability predictions and generalizes to unseen proteins and mutations. Beyond stability prediction, SPURS enables broad applications in protein informatics, including zero-shot identification of functional residues, improved low-N protein fitness prediction, and systematic dissection of stability-pathogenicity for human diseases. Together, these capabilities establish SPURS as a versatile tool for advancing protein stability prediction and protein engineering at scale. Understanding how mutations alter protein stability is essential for biology and disease research. Here, the authors develop SPURS, a model that rewires pre-trained protein generative models to accurately predict stability changes.

CRISPR–Cas systems have revolutionized genome engineering technologies, but type IV CRISPR–Cas systems and their genome engineering potential have been critically underexplored. In this study, we identified a type IV-A3 CRISPR–Cas system from a clinical Klebsiella pneumoniae isolate and characterized its plasmid targeting activity and capacity to suppress chromosomal and plasmid gene expression in E. coli. We revealed the pivotal role of Csf3 (Cas5) and the dispensable roles of Csf1 (Cas8-like) and Csf4 (DinG helicase) subunits in IV-A3 CRISPR–Cas complex formation. The system prevents plasmid propagation via interplay between DinG helicase activity and strategic protospacer positioning relative to plasmid replication and maintenance components. We enabled the IV-A3 CRISPR–Cas system to introduce lethal, sequence-specific double-stranded (ds)DNA breaks in the E. coli chromosome by fusing the nuclease domain of the I-TevI nuclease to the Cas8 N-terminus. Further, we developed a series of base editors, with various editing efficiencies and windows, by fusing the PmCDA1 cytidine deaminase to the Cas8, Cas5, and DinG subunits. Finally, conjugative transfer of the Cas5–PmCDA1 base editor into E. coli deactivated the tryptophan repressor gene, boosting IAA production. Our study provides new insights into type IV-A3 CRISPR–Cas systems and highlights their potential in genome engineering applications.