Abstract
Accelerating Arctic permafrost thaw due to climate warming threatens to release ancient microbial lineages and their functional genetic reservoirs into modern ecosystems. Among the most concerning components of this paleogenome is the uncharacterized environmental resistome. In this study, we combined high-fidelity long-read metagenomic sequencing (Pacific Biosciences HiFi and Oxford Nanopore) with a novel deep learning framework to characterize the antimicrobial resistance determinants within Late Pleistocene Arctic permafrost chronosequences. Long-read metagenomics facilitated the reconstruction of 142 high-quality metagenome-assembled genomes (MAGs) spanning uncultivated phyla, including Actinomycetota, Chloroflexota, and Acidobacteriota, resolving complete operons and mobile genetic elements without fragmentation artifacts. To overcome the limitations of reference-dependent homology searches against modern clinical databases, we deployed a transformer-based protein language model trained on structural and contextual representations of antibiotic resistance genes (ARGs). Our framework identified 87 previously unannotated ARGs exhibiting less than 40% amino acid sequence identity to characterized clinical equivalents, predominantly encoding novel metallo-beta-lactamases, aminoglycoside phosphotransferases, and multidrug efflux pumps. Protein structural modeling via AlphaFold3 and molecular dynamics simulations confirmed that these ancestral enzymes possess structurally intact catalytic clefts capable of binding modern cephalosporins and carbapenems. Our findings reveal that permafrost microbiomes harbor a structurally conserved and functionally diverse intrinsic resistome that predates modern anthropogenic antibiotic selection, underscoring the necessity of combining advanced sequencing modalities with predictive machine learning to evaluate emerging biosafety risks associated with global cryosphere degradation.
