Identification of Bacterial Drug-Resistant Cells by the Convolutional Neural Network in Transmission Electron Microscope Images

Mitsuko Hayashi-Nishino, K. Aoki, Akihiro Kishimoto, Yuna Takeuchi, Aiko Fukushima, Kazushi Uchida, Tomio Echigo, Yasushi Yagi, Mika Hirose, Kenji Iwasaki, Eitaro Shin’ya, Takashi Washio, Chikara Furusawa, Kunihiko Nishino

Introduction

The rapid emergence of antibiotic-resistant bacteria poses a significant global health threat. Traditional detection methods rely on "growth-based" assays, which monitor bacterial proliferation in the presence of drugs. While effective, these methods are time-consuming and offer little insight into the structural consequences of resistance. This research explores whether drug resistance leaves a detectable "morphological signature" on the bacterial cell, allowing for identification using deep learning and high-resolution imaging without antibiotic exposure.

Methodology

The study utilized a combination of advanced imaging and convolutional neural networks (CNNs):

  • Sample Preservation: Researchers employed high-pressure freezing and freeze-substitution to maintain the bacteria's ultrastructure in a near-native state, avoiding the artifacts common in traditional chemical fixation.
  • Transmission Electron Microscopy (TEM): High-resolution images were captured of parental (sensitive) and four independently evolved enoxacin-resistant (ENX) E. coli strains.
  • Deep Learning Model: A batch-normalized AlexNet CNN was trained on approximately one million 512x512 pixel "patches" derived from the larger TEM images.
  • Explainability and Correlation: Grad-CAM (Gradient-weighted Class Activation Mapping) was used to visualize the features the model prioritized. These features were then correlated with gene expression data using Pearson’s coefficients.

Results and Morphological Insights

The CNN achieved a remarkable classification accuracy of 94% (0.94), with sensitivity and specificity scores consistently exceeding 0.90. This high performance suggests that resistance-conferring mutations induce consistent, observable structural changes.

Analysis revealed that enoxacin-resistant cells are morphologically distinct from their parental counterparts. While parental cells maintain a typical rod shape, resistant strains tend to be more spherical. Quantitative morphometric analysis confirmed these changes in cell dimensions, showing shifts in both area and the minor axis of the cells.

Comparison of cell area across parental (P) and enoxacin-resistant (ENX) strains.
Minor axis measurements showing the increased width of resistant strains, contributing to a rounder phenotype.

Grad-CAM visualizations indicated that the AI focused primarily on the cell envelope—specifically the outer membrane and periplasmic space—in resistant cells. These regions often appeared "wavy" with bleb-like structures. In contrast, the model focused on dense granules at the poles of sensitive cells.

Genetic Basis of Structural Changes

The study identified a strong correlation between these structural features and the lpp gene, which encodes a major lipoprotein anchoring the outer membrane. In the resistant ENX strains, lpp was found to be downregulated or mutated. This genetic change likely causes the observed membrane instability and "waviness," providing the morphological signature detected by the CNN.

Discussion and Future Directions

While the results are promising, the study was limited to laboratory-evolved E. coli and a single antibiotic. Future research must determine if these morphological signatures are universal across different species and clinical isolates, which often possess more complex resistance profiles. Additionally, transitioning from patch-based classification to single-cell identification remains a key goal for clinical applicability.

Summary updated

Abstract

The emergence of bacteria that are resistant to antibiotics is common in areas where antibiotics are used widely. The current standard procedure for detecting bacterial drug resistance is based on bacterial growth under antibiotic treatments. Here we describe the morphological changes in enoxacin-resistant <i>Escherichia coli</i> cells and the computational method used to identify these resistant cells in transmission electron microscopy (TEM) images without using antibiotics. Our approach was to create patches from TEM images of enoxacin-sensitive and enoxacin-resistant <i>E</i>. <i>coli</i> strains, use a convolutional neural network for patch classification, and identify the strains on the basis of the classification results. The proposed method was highly accurate in classifying cells, achieving an accuracy rate of 0.94. Using a gradient-weighted class activation mapping to visualize the region of interest, enoxacin-resistant and enoxacin-sensitive cells were characterized by comparing differences in the envelope. Moreover, Pearson's correlation coefficients suggested that four genes, including <i>lpp</i>, the gene encoding the major outer membrane lipoprotein, were strongly associated with the image features of enoxacin-resistant cells.
BibTeX
@inproceedings{hayashinishino2022identification,
  author = {Mitsuko Hayashi-Nishino and K. Aoki and Akihiro Kishimoto and Yuna Takeuchi and Aiko Fukushima and Kazushi Uchida and Tomio Echigo and Yasushi Yagi and Mika Hirose and Kenji Iwasaki and Eitaro Shin’ya and Takashi Washio and Chikara Furusawa and Kunihiko Nishino},
  title = {{Identification of Bacterial Drug-Resistant Cells by the Convolutional Neural Network in Transmission Electron Microscope Images}},
  booktitle = {Frontiers in Microbiology},
  year = {2022},
  doi = {10.3389/fmicb.2022.839718},
}
Figure