Highlights d Cities possess a consistent ''core'' set of non-human microbes d Urban microbiomes echo important features of cities and city-life d Antimicrobial resistance genes are widespread in cities d Cities contain many novel bacterial and viral species

Closing the gap between measurable genetic information and observable traits is a longstanding challenge in genomics. Yet, the prediction of molecular phenotypes from DNA sequences alone remains limited and inaccurate, often driven by the scarcity of annotated data and the inability to transfer learnings between prediction tasks. Here, we present an extensive study of foundation models pre-trained on DNA sequences, named the Nucleotide Transformer, integrating information from 3,202 diverse human genomes, as well as 850 genomes from a wide range of species, including model and non-model organisms. These transformer models yield transferable, context-specific representations of nucleotide sequences, which allow for accurate molecular phenotype prediction even in low-data settings. We show that the representations alone match or outperform specialized methods on 11 of 18 prediction tasks, and up to 15 after fine-tuning. Despite no supervision, the transformer models learnt to focus attention on key genomic elements, including those that regulate gene expression, such as enhancers. Lastly, we demonstrate that utilizing model representations alone can improve the prioritization of functional genetic variants. The training and application of foundational models in genomics explored in this study provide a widely applicable stepping stone to bridge the gap of accurate molecular phenotype prediction from DNA sequence alone.

Proteins have evolved to perform diverse cellular functions, from serving as reaction catalysts to coordinating cellular propagation and development. Frequently, proteins do not exert their full potential as monomers but rather undergo concerted interactions as either homooligomers or with other proteins as hetero-oligomers. The experimental study of such protein complexes and interactions has been arduous. Theoretical structure prediction methods are an attractive alternative. Here, we investigate homo-oligomeric interfaces by tracing residue coevolution via the global statistical direct coupling analysis (DCA). DCA can accurately infer spatial adjacencies between residues. These adjacencies can be included as constraints in structure prediction techniques to predict high-resolution models. By taking advantage of the ongoing exponential growth of sequence databases, we go significantly beyond anecdotal cases of a few protein families and apply DCA to a systematic large-scale study of nearly 2,000 Pfam protein families with sufficient sequence information and structurally resolved homo-oligomeric interfaces. We find that large interfaces are commonly identified by DCA. We further demonstrate that DCA can differentiate between subfamilies with different binding modes within one large Pfam family. Sequence-derived contact information for the subfamilies proves sufficient to assemble accurate structural models of the diverse protein-oligomers. Thus, we provide an approach to investigate oligomerization for arbitrary protein families leading to structural models complementary to often-difficult experimental methods. Combined with ever more abundant sequential data, we anticipate that this study will be instrumental to allow the structural description of many heteroprotein complexes in the future.homo-oligomers | coevolution | direct coupling analysis | protein-protein interactions | big data analysis L ife on the molecular level is orchestrated by the interplay of many different biomolecules. A crucial component for the function is its structure, ranging from small monomers to complex homo-or heteromultimers. The full structural characterization of a biomolecule therefore typically precedes a detailed explanation of its functional mechanism. However, despite the incredible progress of structural characterization methods, many important biomolecules have not been structurally resolved. An intriguing alternative to often involved experimental measurements of 3D structures is taking advantage of the growing wealth of genetic sequential information via sophisticated statistical methods. Direct coupling analysis (DCA) (1, 2) and related tools (3, 4) develop a global model mimicking evolutionary fitness landscapes of protein families (5, 6) and quantify the coevolution of amino acid residue positions (7-9). In the context of protein structure prediction, these models allow extraction of residueresidue contacts from sequence information alone. Such information has proven useful in the prediction of tertiary protein structu...