Detecting Operons in Bacterial Genomes via Visual Representation Learning

Rida Assaf, Fangfang Xia, Rick StevensbioRxiv 2019 · DOI

TL;DR

This paper introduces "Operon Hunter," a novel deep learning approach that treats bacterial operon prediction as an image recognition problem. By converting genomic regions into visual representations, Operon Hunter achieves state-of-the-art performance, particularly in accurately identifying the full extent and boundaries of operons, outperforming existing methods.

What Problem Does This Paper Address?

In bacterial genomes, genes are often organized into functional units called operons, which are crucial for understanding gene function and regulatory networks. Identifying these operons is challenging due to various factors influencing their organization. Traditionally, human experts visually inspect gene neighborhoods across multiple related genomes to identify operons, considering features like gene size, strand direction, inter-genic distance, and evolutionary conservation. While computational tools exist, they often focus on individual gene pairs and struggle to capture the complex interplay of these features, especially when inferring operon boundaries. This paper aims to bridge this gap by developing a method that mimics human visual inspection using deep learning.

Methods

The core innovation of Operon Hunter lies in its visual representation learning approach. Instead of processing tabular data of gene features, the method transforms genomic regions into images. These images are generated using an offline version of the PATRIC Compare Region Viewer, which displays aligned genomes with genes represented as arrows. Key visual features encoded include:

  • Gene Size and Strand Direction: Represented by arrow length and direction.
  • Inter-genic Distance: Scaled within each row.
  • Functional Relatedness: Genes belonging to the same family are colored similarly.
  • Evolutionary Conservation: Achieved by aligning the query genome with related genomes.
  • STRING Score: Encoded in the alpha channel of the image, representing protein-protein association strength.

The model then employs transfer learning, leveraging a pre-trained ResNet18 network (trained on ImageNet) and fine-tuning its top layer on the operon dataset. Data augmentation, specifically horizontal flips, was used to further enrich the training data and mitigate overfitting.

Datasets

  • Training Dataset: Constructed from the "Known Operons" section of the Operon DataBase (ODB). Six genomes with the most labeled operons were selected. Operonic gene pairs were extracted from validated operons, and non-operonic pairs were generated by taking genes adjacent to known operon boundaries. A balanced dataset of 4,306 operonic and 4,290 non-operonic gene pairs was curated.
  • Testing Dataset: Used two well-studied genomes with experimentally validated operons: Escherichia coli (361 operon pairs, 461 non-operon pairs) and Bacillus subtilis (369 operon pairs, 302 non-operon pairs). These were sourced from RegulonDB and DBTBS, respectively, and cross-checked with ODB. A leave-one-out approach was used, meaning genomes used for testing were excluded from training.

Key Insights and Ideas

  • Visualizing Genomics for Deep Learning: The paper demonstrates that complex biological relationships, like operon structures, can be effectively captured by transforming them into visual representations, making them amenable to powerful image recognition models.
  • Mimicking Human Expertise: By replicating the visual inspection process used by human experts, the model learns to integrate diverse features (gene-level, functional, and evolutionary) in a holistic manner.
  • Interpretable AI: The use of Grad-Cam allows for visualization of the network's decision-making process, highlighting which visual features (e.g., strand direction, inter-genic distance, neighborhood conservation) most influence its predictions. This provides a crucial link for expert validation.
  • Transfer Learning's Power: Leveraging pre-trained models significantly boosts performance, especially with limited specialized datasets, showcasing the broad applicability of vision models to biological problems.

Key Results

  • Superior Operon Prediction: Operon Hunter achieved the highest accuracy in gene-pair predictions, with an F1 score of 0.921 and an ROC AUC of 0.97.
  • Excellent Full Operon Prediction: The method excelled in predicting full operons, correctly identifying 85% of them, significantly outperforming ProOpDB (62%) and Door (61%). This indicates a strong ability to delineate operon boundaries.
  • Stable Performance: The model showed stable performance across different metrics and across the two tested genomes (E. coli and B. subtilis), suggesting good generalization capabilities.
  • Visual Feature Importance: Heatmaps revealed that for operonic predictions, the model focuses on the immediate vicinity of the gene pair (inter-genic distance, strand direction), while for non-operonic predictions, it considers a broader context, including non-conserved alignments and low functional scores.

Limitations and Future Work

  • Dataset Size: While transfer learning and data augmentation were used, the dataset of validated operons is still relatively limited, posing a risk of overfitting.
  • Comparison Challenges: The authors note difficulties in directly comparing their results with other tools due to differences in how operons (especially single-gene operons) are defined and how training/testing datasets are constructed.
  • Future Directions: The authors suggest that this visual representation approach could be extended to other comparative genomics problems where ground truth datasets are limited. They also hypothesize that this trend of using image-based representations will continue to grow in scientific domains.

Summary updated

Abstract

ABSTRACT Contiguous genes in prokaryotes are often arranged into operons. Detecting operons plays a critical role in inferring gene functionality and regulatory networks. Human experts annotate operons by visually inspecting gene neighborhoods across pileups of related genomes. These visual representations capture the inter-genic distance, strand direction, gene size, functional relatedness, and gene neighborhood conservation, which are the most prominent operon features mentioned in the literature. By studying these features, an expert can then decide whether a genomic region is part of an operon. We propose a deep learning based method named Operon Hunter that uses visual representations of genomic fragments to make operon predictions. Using transfer learning and data augmentation techniques facilitates leveraging the powerful neural networks trained on image datasets by re-training them on a more limited dataset of extensively validated operons. Our method outperforms the previously reported state-of-the-art tools, especially when it comes to predicting full operons and their boundaries accurately. Furthermore, our approach makes it possible to visually identify the features influencing the network’s decisions to be subsequently cross-checked by human experts.
BibTeX
@misc{assaf2019detecting,
  author = {Rida Assaf and Fangfang Xia and Rick Stevens},
  title = {{Detecting Operons in Bacterial Genomes via Visual Representation Learning}},
  year = {2019},
  doi = {10.1101/860221},
  howpublished = {bioRxiv},
}
Figure