Machine Learning Suggests That Small Size Helps Broaden Plasmid Host Range

Abstract

Plasmids mediate gene exchange across taxonomic barriers through conjugation, shaping bacterial evolution for billions of years. While plasmid mobility can be harnessed for genetic engineering and drug-delivery applications, rapid plasmid-mediated spread of resistance genes has rendered most clinical antibiotics useless. To solve this urgent and growing problem, we must understand how plasmids spread across bacterial communities. Here, we applied machine-learning models to identify features that are important for extending the plasmid host range. We assembled an up-to-date dataset of more than thirty thousand bacterial plasmids, separated them into 1125 clusters, and assigned each cluster a distribution possibility score, taking into account the host distribution of each taxonomic rank and the sampling bias of the existing sequencing data. Using this score and an optimized plasmid feature pool, we built a model stack consisting of DecisionTreeRegressor, EvoTreeRegressor, and LGBMRegressor as base models and LinearRegressor as a meta-learner. Our mathematical modeling revealed that sequence brevity is the most important determinant for plasmid spread, followed by P-loop NTPases, mobility factors, and β-lactamases. Ours and other recent results suggest that small plasmids may broaden their range by evading host defenses and using alternative modes of transfer instead of autonomous conjugation.

Keywords: antibiotic resistance genes; clustering; conjugation; horizontal gene transfer; machine learning; plasmid host range; small-size plasmid.

Figure 1

Overview of the dataset. Plasmid size (A) and GC content (B) distribution. Mean and median values are indicated by blue and green dots, respectively. The cutoffs for categorical terms are indicated with orange dots. (C) Distribution of plasmid-borne proteins among COGs (shown on the right). The pie chart shows the proportion of proteins in COGs. The bar chart shows the fraction of proteins assigned to a given COG; fractions of the Z, A, and B categories are indicated by numbers. The three most prevalent COG categories are highlighted in the table.

Figure 2

Plasmid clustering. Every node represents a plasmid. The network includes 30,464 plasmids assigned into 1125 clusters (≥3 members). The nodes are colored according to the three plasmid mobility types: Orange, conjugative. Blue, mobilizable. Green, non-mobilizable.

Figure 3

Defining the plasmid distribution possibility. (A) The sum of distribution possibility (P_sum). Meanings of the P_sum for distribution levels are indicated. (B) Phyla distribution for clusters with P_sum > 4.7. Only clusters with more than 100 plasmids were considered. NPn_tax: for a given taxon, the plasmid number is normalized by the average plasmid number per cluster.

Figure 4

Performance of the final model stack. The dataset was split 100 times randomly into training (A) and testing (B) dataset pairs. The model stack was trained on the training dataset, and then tested on paired unseen testing datasets. MSE, mean squared error. R2, coefficient of determination. MSE (P_sum > 1), MSE for those clusters with P_sum > 1. Values are reported as mean ± SD (n = 100). The data points on the very right of the training dataset were synthesized by SMOTE.

Figure 5

Feature importance analysis. (A) An overview of feature types. The outer rim of the diagram represents features grouped into 12 main feature types. The mobility of feature types are features involved in plasmid mobility, which is not the same as the three plasmid mobility types. Features that co-occur in different types are linked by edges. (B) Importance index calculated by drop-column analysis. Positive values indicate that the feature removal worsens model stack performance, whereas negative values indicate features whose removal improves model performance. An importance index of 0 indicates that feature removal has no effect. (C) Model stack performance on testing dataset after removal of “smallest”. Values are reported as mean ± SD (n = 100). MSE was calculated by using 100 randomly split training and testing dataset pairs as in Figure 4. (D) Feature type distribution analysis from randomly sampled seven-feature combinations. All sampled combinations, the 11,000 combinations. Selected combinations, the 33 selected combinations. (E) Detailed analysis of the backbone features in all vs. selected combinations. (F) The mobility types distribution in plasmids of different sizes.