Automated Analysis of Fluorescence Microscopy Images to Identify Protein-Protein Interactions

Sankar Venkataraman, Mitchel J. Doktycz, Hairong Qi, Jennifer Lynn Morrell-Falvey

TL;DR

This paper presents an automated image analysis system for identifying protein-protein interactions in bacterial cells using fluorescence microscopy. By developing a robust algorithm to analyze localization patterns, the system achieves high accuracy, significantly speeding up the analysis of large biological datasets and overcoming manual analysis limitations.

What Problem Does This Paper Address?

The postgenomic era has led to an explosion of data, particularly from high-throughput imaging studies in proteomics. Analyzing these vast datasets, especially those involving fluorescence microscopy, is a major bottleneck. Manual analysis is time-consuming, cumbersome, and prone to subjective interpretation. While image acquisition has been automated, the analysis of biological images, particularly for complex phenomena like protein localization, still heavily relies on manual scoring. This paper addresses the need for an automated system to efficiently and accurately analyze fluorescence microscopy images to identify protein-protein interactions, a crucial step in understanding biological networks.

Methods

The core of this work is an automated image analysis pipeline designed to process both Differential Interference Contrast (DIC) and fluorescence microscopy images of E. coli cells.

  • Image Preprocessing and Segmentation:
    • DIC Images: The pipeline enhances cell boundaries by applying a bottom-hat morphological operation, followed by a second-order Butterworth high-pass filter and adaptive histogram equalization. This process aims to create clear, distinct cell outlines. A Canny edge detector is then used for segmentation, with a parameter ($\sigma$) optimized for edge detection.
  • Fluorescence Images: Background noise is reduced by using the cell boundaries identified from the DIC images. An automated global thresholding algorithm (Otsu's method) is applied, with a modification to include weak signals. Morphological operations (opening, closing, median filter) are then used to refine the resulting binary image.
  • Feature Extraction: A set of key features are extracted from the processed images, including:
  • Number of cells in an image.
  • Area of individual cells.
  • Area of GFP localization sites within a cell.
  • Diameter of individual cells.
  • Center of gravity (COG) of cells and GFP localization sites.
  • Distance of GFP localization sites from the cell's COG.
  • Number of GFP localization sites within each cell.
  • Pattern Recognition: The system uses a decision rule based on established experimental findings: a positive interaction is indicated by the presence of two (at cell poles) or three (poles and medial region) GFP-fusion protein localization sites. The algorithm quantifies localization sites and their positions, using the extracted features to classify interactions. A key aspect is distinguishing true localization from "inclusion bodies," which can mimic positive interactions. This is addressed by analyzing images before induction of the DivIVA-fusion protein.

Datasets

The study utilizes a dataset of 16 DIC and corresponding fluorescence images. These images were captured from two experiments involving E. coli cells expressing GFP-fusion proteins. The dataset represents approximately 390 cells in total, acquired at different magnifications and from different samples to ensure robustness. The authors do not introduce a new benchmark dataset but rather evaluate their algorithm against expert scoring of this existing dataset.

Key Insights and Ideas

  • Automated analysis is feasible and accurate: The paper demonstrates that automated analysis of fluorescence microscopy images for protein-protein interactions is not only possible but can achieve high accuracy, rivaling expert scorers.
  • DIC images are crucial for cell segmentation: Leveraging DIC images for precise cell boundary identification is a critical step that enables accurate segmentation of fluorescence signals within individual cells.
  • Distinguishing inclusion bodies is vital: The presence of inclusion bodies, a common artifact of protein overexpression, can lead to false positives. The paper highlights the importance of pre-induction imaging to identify and exclude samples with inclusion bodies.
  • Feature selection is key: A carefully chosen set of features, particularly those related to the number and location of GFP localization sites, are sufficient for robust classification of protein interactions.
  • Parameter optimization is necessary: The performance of the algorithm is sensitive to certain parameters, such as the standard deviation ($\sigma$) in the Canny edge detector and the thresholding value for fluorescence segmentation. The paper investigates the impact of these parameters.

Key Results

  • The automated system achieved 100% accuracy in detecting positive interactions for the dataset of 16 images, based on the criterion of identifying at least 50% positive cells in an image.
  • The algorithm demonstrated an average sensitivity of 86% for identifying individual cells.
  • The average sensitivity for determining the presence of positive localization patterns within identified cells was approximately 84.39%.
  • The system achieved a mean specificity of about 0.9995 for cell identification and 0.9989 for interaction identification, indicating a very low rate of false positives.
  • The choice of $\sigma = 0.85$ in the Canny edge detector provided the best robustness for cell identification, yielding an average sensitivity of 86%.
  • A threshold value of one-third of Otsu's method output provided the highest sensitivity for positive interaction identification.

Limitations and Future Work

  • Inclusion Bodies: While the system attempts to mitigate the issue of inclusion bodies by pre-induction analysis, it acknowledges that this problem cannot be completely eliminated and can still lead to false positives.
  • Experimental Factors: Low sensitivities in some cases were attributed to experimental and biological factors such as the focal plane of the collected image and plasmid loss.
  • Adaptability: While the current system is tailored for the DivIVA-based assay, the authors suggest it could be adapted to a wider range of experimental studies involving different fluorescent labels or imaging modalities.
  • Dataset Size: The evaluation was performed on a dataset of 16 images. While this provided robust results, larger and more diverse datasets could further validate the system.
  • Future Directions: The authors envision the system being employed to reduce image dataset sizes by selecting images with desired features, such as positive interactions or specific localization patterns. The system's adaptability to other experimental setups is also a promising avenue.

Summary updated

BibTeX
@article{venkataraman2006automated,
  author = {Sankar Venkataraman and Mitchel J. Doktycz and Hairong Qi and Jennifer Lynn Morrell-Falvey},
  title = {{Automated Analysis of Fluorescence Microscopy Images to Identify Protein-Protein Interactions}},
  journal = {Int. J. Biomed. Imaging},
  year = {2006},
  doi = {10.1155/ijbi/2006/69851},
}
Figure