[P004] Poster

1:00 PM - 1:50 PM JST(4:00 AM - 4:50 AM UTC)

[P02-228] Prediction of plasmid destination using nucleotide sequence and machine learning

○Maho TOKUDA¹, Masaki Shintani^1,2 (1. Shizuoka University (Japan), 2. RIKEN BRC-JCM (Japan))

Keywords:

plasmid,conjugative transfer,host range,machine learning

Plasmids play a crucial role in microbial ecology by mediating horizontal gene transfer (HGT) to spread functional genes, such as antibiotic resistance and metabolic genes, among diverse bacteria. This process not only affects microbial community structures and ecosystem dynamics but also supports environmental biotechnological processes, including pollutant degradation, nutrient cycling, and adaptation of microbial consortia to anthropogenic environments. To better understand and control these processes, it is essential to identify bacteria capable of acquiring plasmids (here defined as “plasmid destinations”). However, experimentally determining plasmid destinations across diverse bacteria remains challenging, emphasizing the need for bioinformatics approaches that can accurately predict them from nucleotide sequences. In this study, we aimed to develop a prediction method for plasmid destination using machine learning. A total of 1,778 conjugation assays were performed using 14 model plasmids (seven IncP/P-1, five PromA, one pSN1216-29-like plasmid and one unassigned group plasmid) with Pseudomonas as the donor and 127 recipients (covering five phyla with available complete genomes sequences). Plasmid transfer was detected by flow cytometry using Transfer was observed in 856 plasmid-recipient combinations (destination positive), whereas no transfer was detected in 922 combinations (destination negative). Comparison of GC content and k-mer compositions showed that exhibited higher plasmid-chromosome similarity than destination-negative ones, consistent with previous findings regarding plasmid evolutionary hosts¹⁾. Based on these findings, we constructed prediction models using Random Forest with features (1) derived from k-mer composition similarity and (2) embeddings generated by the DNA language model DNABERT.We then applied the models to predict the transferability of the unclassified plasmid pYKAS102 to the same 127 recipients. Furthermore, when applied to IncP-1 η, ρ, and μ subgroup plasmids, previously thought to have narrower host ranges restricted to Enterobacteriaceae, the models predicted significantly fewer positive destinations compared to well-studied IncP-1 plasmids such as RK2²⁾. Interestingly, in addition to , Acinetobacter species were also predicted to be destination positive. Conjugation assays confirmed that these plasmids could indeed be transferred to Acinetobacter, supporting the validity of the prediction model. Overall, our approach provides a framework for predicting plasmid destinations directly from nucleotide sequences, facilitating a foundation for understanding of HGT in microbial communities. Future efforts will focus on enhancing prediction accuracy by integrating additional sequence-derived features. 1)Suzuki H et al., 2010 J Bacteriol.;192(22):6045-55. 2)Hayakawa M et al., 2022 Appl Environ Microbiol 88(18):e0111422