Genomic-context models have shown that contextualizing protein embeddings can improve predictions of microbial function. We observe that a gene's protein embedding is partially predictable from the embeddings of neighboring genes in the genome, motivating high-ratio masking that requires reconstruction from sparser genomic context. Genomic modeling must also distinguish biological relationships from assembly conventions, since the ordering of separate contigs does not imply adjacency and strand labels depend on the reference orientation. We therefore introduce VenusGenome, a gene-resolution masked autoencoder that combines visible-only encoding with genome-aware attention incorporating contig boundaries, relative strand orientations, and gene positions. We pretrain this model on a curated corpus of over 1.7 million microbial genome assemblies from isolate and metagenomic sources, using frozen protein embeddings as gene-level inputs. Across genome-level phenotype prediction and gene-level tasks, the resulting frozen representations achieve the highest mean performance among the evaluated models on nine of eleven primary metrics. An exploratory gene-input perturbation analysis further links representation sensitivity to experimentally measured fitness defects. Ablations support the contribution of genome-aware attention and show the benefits of masked autoencoding over BERT-style masked modeling, with further gains from higher masking ratios.

Glucagon-like peptide-1 (GLP-1) receptor agonists and GLP-1 analogs are widely used for type 2 diabetes (T2D) and weight management, but current approaches can be limited by manufacturing complexity, cost, formulation requirements, and the need for repeated administration. Living microbial delivery systems offer a potential strategy for sustained gastrointestinal production of therapeutic peptides, but achieving stable expression, intestinal survival, and biologically relevant systemic exposure remains challenging. Here, we developed an engineered Bacillus subtilis spore-based platform for the gastrointestinal delivery of pentameric GLP-1. By leveraging the stress resistance, storage stability, and genetic tractability of B. subtilis, we generated a chromosomally integrated strain that secretes a multimeric GLP-1 construct designed to be processed in the intestine into monomeric GLP-1 peptides. Oral administration of engineered spores resulted in a mean serum GLP-1-related peptide concentration of 20.61 and 23.96 nM on days 3 and 5, respectively, in the first trial and 18.72 and 29.27 nM on days 3 and 7, respectively, in an independent second trial, with minimal detectable signal in green fluorescent protein (GFP) spore-treated controls. Treatment was also associated with preliminary glucose-lowering activity in a fasting-refeeding test. In a longer-term study, engineered spore administration was associated with reduced body weight and adiposity in mice with pre-existing diet-induced obesity, with semaglutide included as a positive pharmacologic benchmark. To improve its translational application, we further generated a marker-free strain that retained probiotic-relevant features of the parental strain, including gastrointestinal stress tolerance, epithelial cell adhesion, and antibiotic susceptibility. These findings provide proof of concept that engineered spore-forming bacteria may serve as a gastrointestinal peptide delivery platform and support further evaluation of this approach in disease-relevant models of metabolic dysfunction.

Defense-associated reverse transcriptases (DRTs) mediate diverse antiphage responses, yet how they sense phage infection and couple unconventional DNA synthesis to antiviral effector activation remains poorly understood. Here, we characterize DRT4 as a dual-sensor immune system that integrates two distinct phage-derived signals. Cryo-EM analysis reveals that DRT4 assembles into a dimer-of-trimers hexamer, with each protomer catalyzing protein-primed, template-independent synthesis of a covalently linked ssDNA product. DRT4 displays a strong preference for purine nucleotides, determined by base-specific interactions within the polymerase active site. Using phage escape genetics, we identify two independent activators of DRT4: the phage ribonucleotide reductase NrdAB, which increase intracellular dNTP levels, and the SSB-like protein 0087, which directly binds the DRT4-generated ssDNA. Activated DRT4 exhibits dNTP-dependent RNase activity, linking phage sensing to RNA cleavage and abortive infection. Together, our findings show DRT4-generated ssDNA as a signaling intermediate that integrates phage-induced metabolic changes with direct recognition of viral replication proteins, revealing how protein-primed DNA synthesis can be repurposed for infection sensing and antiviral effector activation.

Genomic foundation models (GFMs) have transformed our ability to learn the evolutionary language encoded in genomes, yet their potential for programmable regulatory DNA engineering remains underexplored. Regulatory function depends on combinatorial interactions among distributed sequence elements, motivating generation with bidirectional context. Here, we introduce DAPE-PD (Diffusion Adaptation of Pretrained Evo for Promoter Design), which adapts the autoregressive Evo model into a bidirectional masked diffusion model without foundation-scale retraining. DAPE-PD combines pretrained genomic representations with iterative unmasking to generate regulatory DNA sequences. Following genus-specific fine-tuning on genome-derived intergenic sequences, DAPE-PD generated sequences that retained genus-associated compositional features despite low similarity to the fine-tuning datasets. Experimental validation identified functional de novo promoters, while in silico prioritization recovered active designs from small experimental candidate sets across three non-model gut commensals. This framework connects genomic pretraining with diffusion-based generation, extending GFMs from sequence understanding toward functional cis-regulatory DNA engineering in living cells.