RNA inverse folding designs nucleotide sequences expected to adopt a prescribed secondary structure. Search-based solvers can optimize folding-model objectives effectively, but difficult targets can require extensive sampling, and structural optimization alone does not explicitly preserve the sequence distributions or conserved motifs of natural RNA families.  We developed RIFT-VAE, a Transformer-based conditional variational autoencoder that receives a context-free grammar parse-tree representation of a target secondary structure and generates nucleotide labels on the corresponding tree. The framework combines progressively richer grammar rules, self-refinement learning from generated structure-sequence pairs, and cross-entropy-method optimization in the learned latent space. We evaluated RNAfold minimum-free-energy agreement on an RNAcentral-derived test set and the EteRNA100 benchmark, compared the method with four search-based solvers under matched total time budgets, and examined GC-content control and covariance-model family annotation. The complete pipeline achieved RNAfold-Correct/RNAfold-MCC values of 0.833/0.994 on the RNAcentral-derived test set and 0.760/0.977 on EteRNA100. Latent-space optimization accounted for the largest increase in exact structural recovery. Under a 3,600-s total budget on EteRNA100, sequences generated by RIFT-VAE improved the exact-match rate of every tested downstream search method when used as warm starts; the largest change was observed for RNAInverse (Correct, 0.297 to 0.803; MCC, 0.505 to 0.985). The pretrained model also produced sequences with measurable correct-family covariance-model hits and supported explicit GC-content conditioning. RIFT-VAE is best interpreted as a hybrid generative-search framework: pretraining supplies a structure- and family-informed proposal distribution, whereas latent optimization concentrates evaluations in high-scoring regions. The reported structural scores are specific to RNAfold minimum-free-energy validation and do not establish biochemical function. Orthogonal folding predictors, stricter homology-controlled splits, diversity-aware evaluation, architecture-matched dot-bracket ablations, and experimental assays remain priorities for validation.

Cross-feeding mutualisms, in which partner species exchange essential metabolites, are ubiquitous in microbial communities. In spatially structured environments, motility can improve access to partner-produced resources but also impose metabolic costs and displace cells from nutrient-rich regions, so its net benefit depends on the spatial dynamics of the interaction. Here, we combine competition experiments in a cross-feeding mutualism between Escherichia coli and Salmonella enterica with a spatially explicit consumer-resource model to determine what drives selection on motility. Spatial structure imposes asymmetric selection between partners i.e. S. enterica benefits from motility regardless of partner motility, whereas selection on E. coli switches from favorable to unfavorable depending on whether its partner can move. Competition in well-mixed culture suggests that this reversal reflects the loss of a spatial benefit rather than an increased cost. Our model attributes the asymmetry to three interacting factors: the ratio of metabolite production to consumption which sets whether the cross-fed resource is scarce or abundant; the number of growth-limiting resources which determines whether an alternative gradient can rescue the benefit of motility; and partner motility and growth rate, which shape where metabolites are produced. When a metabolite is scarce, motile cells gain by dispersing into regions it has reached but not yet been depleted from. When it accumulates, this gradient is eroded, and the motility costs offset any benefit it provides. Selection on motility therefore depends on the metabolic structure of the interaction and the spatial behavior of partners.

Biosynthetic gene clusters (BGCs) are responsible the biosynthesis of many natural products, including a multitude of effective therapeutics and their precursors. Advances in genomic data collection as well as computational techniques have made it possible to identify BGCs at scale. However, accurately determining the types of BGC-encoded products from genomic content remains elusive.  Here, we introduce BGCat (BGC annotation tool), a machine learning method for fine-grained structural classification of BGC-encoded products, leveraging the NPClassifier natural product nomenclature. Our method leverages a pre-trained protein language model for creating meaningful gene representations and a deep neural network for class label prediction. We show the method outperforms state-of-the-art approaches in coarse-grained product classification and is effective for detailed classification. We implement a clustering-based augmentation strategy for BGC-product relationships, addressing a crucial gap in the available datasets. We then introduce the concept of product class profiles (PCPs) of gene cluster families (GCFs), associating each GCF with a probabilisitc distribution of product types and offering a new perspective on GCF functions. Lastly, we use BGCat to provide new product class labels for over 100k BGCs in antiSMASH DB that presently have minimal information about their products.

Multicellular organization enables biological functions to be distributed among specialized cells and coordinated through intercellular communication. Integrating this organizational principle with genome replication would link functional division of labor to the propagation of genetic information. Here, we show that genomic RNAs with complementary functions can cooperatively replicate across communicating artificial compartments. We constructed multicell-like colonies from all-aqueous droplets formed by phase separation of two incompatible polymers and stabilized at their interfaces by liposomes and amyloid-like proteins. The droplets assembled spontaneously while remaining permeable to protein-sized macromolecules. Two genomic RNAs encoding a replication enzyme and a metabolic enzyme were distributed in distinct colony-forming droplets and cooperatively replicated through cell-free translation and reciprocal molecular communication. These findings establish that genome replication can be collectively supported by communicating artificial compartments and provide a route toward multicell-like systems that coordinate spatially distributed genetic functions.

Biological AI learns primarily from existing observations, but many questions cannot be answered from available data alone. Here we show that AI can instead acquire knowledge by acting directly on biological systems and learning from the consequences. We developed a closed-loop framework in which autonomous agents design protein variants, construct and characterize them in a robotic laboratory, learn from the resulting experimental feedback, and decide what experiments to perform next. We then allowed the system to operate continuously and without human intervention for approximately one month, during which multiple agents independently explored protein sequence space while learning from shared experimental experience. Applied to glycoside hydrolases, the agents discovered enzymes with substantially altered substrate specificity toward non-native sugars and progressively learned the structure of the underlying sequence-function landscape. The resulting experimental experience also revealed determinants of substrate specificity and protein expression that were not specified as learning objectives. These results demonstrate that AI can autonomously interact with biology over extended periods to acquire knowledge through experience, establishing a framework for biological discovery driven by continuous experimental interaction.

The dominant paradigm for temperature robustness in biomolecular circuits is for the parameters to be tuned to have matching temperature dependencies so that their overall effect cancels out. This contrasts with the robustness due to circuit structure, typically operative in circuits where robustness to a single input parameter is desired. The importance of the circuit structure in temperature robustness is generally unclear. We addressed this issue in a benchmark negative feedback circuit using a combination of theoretical modelling and experimental measurements. We found that the response to a temperature perturbation in a model of negative feedback was qualitatively different from the response in a model without feedback. We experimentally measured the response of the negative feedback circuit to a temperature perturbation and found that it was smaller than that of the circuit without feedback, in line with the theoretical finding. We confirmed this theoretical prediction experimentally. The initial response of the negative feedback circuit, paradoxically, was larger than the circuit without feedback. The resolution of this paradox was in accounting for the faster dynamics in the negative feedback circuit. These results show a simple design principle of temperature robustness that can operate in a widespread circuit motif and may also apply to other perturbations which, like temperature, affect multiple parameters simultaneously.

IS110 has emerged as a powerful genome-editing tool because it is the smallest RNA-guided system capable of diverse programmable insertions. Naturally existing elements are conventionally modeled as compact ~1.5-kb systems comprising a single transposase and a bridge RNA (bRNA). Using high-throughput junction mapping together with large-scale comparative genomics, we redefined the in vivo structural boundaries, growth, and mobilization of IS110 elements. We uncovered a previously unrecognized size continuum extending to ~100 kb, driven by progressive local expansion, with expanded loci being widespread across bacterial genomes. Experiments confirmed activity of natural IS110s both well below and above the size range of previously characterized elements. The large systems preferentially accumulated adaptive cargo, including antimicrobial resistance determinants and heavy-metal detoxification systems, and were strongly enriched for plasmid-derived DNA. Boundary configurations at expanded loci and the range of partial excision intermediates they produce both indicate flexible sequence recognition by IS110, most commonly through half-matches between the bRNA and complementary DNA sequence. This sequence tolerance allows loci to expand with diverse cargo. Together, these findings redefine IS110 from a compact insertion sequence into a dynamic platform that disseminates adaptive cargos.

Predicting protein–RNA binding affinity is crucial for understanding cellular regulation and advancing RNA-targeted drug discovery. However, this task remains challenging due to structural complexity, limited labeled data, and insufficient modeling of fine-grained interactions.  We propose M2-PRNet, a multi-scale and multi-modal framework that integrates atom-level graphs, residue-level graphs, and tri-view molecular representations to capture complementary structural information. A cross-scale contrastive learning objective is introduced to align representations across different structural resolutions of the same complex. Under a clustering-based five-fold cross-validation setting on benchmark datasets, M2-PRNet achieves state-of-the-art performance. To further assess generalization under reduced sequence homology, we construct homology-aware RNA-cold, protein-cold, and dual-cold evaluations under a stricter 40% sequence identity threshold, where M2-PRNet maintains competitive performance. To account for conformational flexibility, we evaluate the model on MD150-1ns and an extended MD75-10ns subset, demonstrating stable performance under MD-derived structural perturbations. In addition, representative case studies suggest that M2-PRNet can highlight relevant RNA-binding regions and support preliminary discrimination between strong and weak binders when plausible complex structures are available. These results demonstrate the effectiveness of integrating multi-scale and multi-modal representations with cross-scale alignment for protein–RNA affinity prediction.

The problem of how protein sequences translate into defined functions remains largely unsolved despite decades of progress. New methods to efficiently explore protein sequence space will help to shed light on these sequence-function relationships, particularly for complex protein function. Here, we describe an approach to create novel, functional proteins through the integration of deep mutational scanning, structural analysis, and evolutionary mining within prompts for a generative protein language model (PLM). We demonstrate the utility of this approach with the generation of novel compact RNA-guided nucleases. This approach is highly efficient, resulting in active nucleases with ~40% sequence divergence relative to natural proteins and activity equivalent to or exceeding by up to ~3X that of other compact nucleases at multiple endogenous loci in human cells. The approach described here is rapidly deployable and produces new sequences that will serve as scaffolds for further exploration of complex protein functionality, as well as substrates for novel genome engineering applications.

Codon optimization uses synonymous sequence changes to improve the expression and therapeutic performance of nucleic acid-based medicines. Masked language models (MLMs) have recently been proposed as alternatives to traditional, frequency-based codon optimization approaches, yet whether they offer a meaningful advantage over such simpler methods remains unclear. Here we benchmark three prominent MLMs --- CaLM, EnCodon and CodonTransformer --- across backtranslation fidelity, sequence generation and nine molecular phenotype prediction tasks, and experimentally evaluate model-designed sequences using a secreted embryonic alkaline phosphatase (SEAP) reporter. The models differed markedly in amino-acid fidelity and generated distinct synonymous sequence variants. However, no single model performed best across all benchmark tasks and simple sequence features remained competitive in several settings. Our interpretability analysis revealed that the models integrate a large window of codon context for making predictions, as opposed to frequency-based approaches. Our in vitro data showed that MLM-designed variants outperformed conventional and commercial vendor-derived sequences in both transient and stably integrated expression, supporting the models' ability to capture translational context beyond codon frequency. Together, our results establish MLMs as effective and complementary tools for codon optimization and suggest that sampling across multiple models may improve the likelihood of identifying high-performing therapeutic sequences.