Candidate gene prioritization using graph embedding

Quan Do, Pierre Larmande

TL;DR

This paper tackles the challenge of identifying candidate genes associated with specific traits in rice by leveraging knowledge graphs and advanced embedding techniques. They introduce a new dataset and demonstrate that Convolutional Neural Networks (CNNs) applied to knowledge graphs can effectively rank genes, outperforming traditional methods and enabling downstream analysis like gene clustering and similarity search.

What Problem Does This Paper Address?

Identifying genes linked to specific phenotypes or diseases is crucial for agricultural advancements, especially in crops like rice, a staple food. Genome-Wide Association Studies (GWAS) often yield hundreds or thousands of potential candidate genes, making experimental verification costly and time-consuming. Integrating and analyzing the vast, fragmented data available on gene functions, traits, and interactions is a significant hurdle. Existing computational methods for gene prioritization are increasing, but few effectively handle the complexity of gene-phenotype relationships. This paper aims to bridge this gap by using knowledge graphs and deep learning to represent and analyze these intricate relationships.

Methods

The core of this work lies in applying knowledge graph embedding techniques to represent genes and their attributes as vectors. They explore two main embedding models:

  • Translating Embedding (TransE): A foundational model that represents entities (genes, attributes) and relations as vectors, aiming to capture the relationship $h + r \approx t$ (head + relation ≈ tail). TransE is used as a pre-training step to initialize gene embeddings.
  • Convolutional Knowledge Base (ConvKB): This model extends TransE by using Convolutional Neural Networks (CNNs) to learn more complex, non-linear relationships within triples (head, relation, tail). CNNs are adept at capturing local patterns and can potentially learn richer representations than purely linear models. ConvKB is designed to capture relationships between entities and their attributes, and the paper suggests it can generalize beyond simple one-to-one relationships.

After embedding, they use unsupervised learning techniques:

  • K-Means Clustering: To group similar genes based on their learned vector representations.
  • K-Nearest Neighbors (KNN): To find genes with similar attributes or rank genes based on attribute similarity.

Datasets

The authors introduce a new dataset named OsGenePrio specifically for rice gene information. This dataset was constructed by integrating data from several open-access rice databases, including:

  • Oryzabase
  • RAP-DB (Rice Annotation Project Database)
  • Rice SNP-Seek Database
  • Funricegene (covering gene keywords, family info, and general info)
  • Gramene
  • MSU Rice Genome Annotation Project (RGAP)
  • Predicted Rice Interactome Network (PRIN)
  • UniProt

They also mention using the PyRice package for data collection and filtering. The dataset was pre-processed to standardize attributes and remove those with very low or high frequencies to improve model performance. The final dataset consists of 12 relations and 33,268 entities, split into training and validation sets.

Key Insights and Ideas

  • CNNs for Knowledge Graphs: The paper demonstrates that CNNs, typically used in computer vision and NLP, can be effectively adapted for knowledge graph completion and analysis in bioinformatics. This suggests a broader applicability of CNNs beyond their traditional domains.
  • Richer Gene Representations: ConvKB's ability to capture non-linear relationships allows for potentially more nuanced and informative vector representations of genes and their attributes compared to simpler models like TransE.
  • Unsupervised Analysis: The learned gene embeddings can be directly used for downstream tasks like clustering and similarity search, providing valuable biological insights without requiring additional supervised labels. This is crucial for exploratory data analysis.
  • Dataset Creation: The creation of a dedicated, integrated dataset (OsGenePrio) is a significant contribution, addressing the fragmentation of existing rice gene information.

Key Results

The paper evaluates the models using link prediction metrics (Mean Rank, Mean Reciprocal Rank, Hits@). While specific comparative numbers against other state-of-the-art methods are not explicitly detailed for all metrics, the results indicate that the ConvKB model, when properly tuned, achieves strong performance. For instance, they report a Mean Reciprocal Rank (MRR) of 0.7 and Hits@10 of 0.914 on the validation set using the "Filtered" protocol with their best parameter configuration. The K-Means clustering analysis suggests that an optimal number of clusters for their dataset lies around 75-100, indicating a reasonable ability to group genes. The K-nearest neighbor results show promising similarity findings between gene searches and the proposed model.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • ConvKB Drawbacks: They note that ConvKB, while effective, has limitations. The paper doesn't fully elaborate on these, but it's implied that further enhancements are possible.
  • Data Noise and Missing Information: The dataset, despite pre-processing, may still contain noise and missing gene information, which can hinder training and prediction accuracy. Researchers in biology may need to confirm predictions.
  • Model Enhancements: The current model primarily uses TransE and a basic CNN architecture. Future work could explore more advanced knowledge graph embedding models and CNN architectures for improved performance and efficiency.
  • Expanding PyRice: Integrating the ConvKB model into the PyRice tool could make gene search more efficient and meaningful, moving beyond simple text searches. Expanding PyRice to include more databases is also a goal.
  • Broader Database Integration: The authors suggest expanding the model to incorporate more rice gene databases.

Summary updated

Abstract

Abstract Candidate genes prioritization allows to rank among a large number of genes, those that are strongly associated with a phenotype or a disease. Due to the important amount of data that needs to be integrate and analyse, gene-to-phenotype association is still a challenging task. In this paper, we evaluated a knowledge graph approach combined with embedding methods to overcome these challenges. We first introduced a dataset of rice genes created from several open-access databases. Then, we used the Translating Embedding model and Convolution Knowledge Base model, to vectorize gene information. Finally, we evaluated the results using link prediction performance and vectors representation using some unsupervised learning techniques.
BibTeX
@inproceedings{do2020candidate,
  author = {Quan Do and Pierre Larmande},
  title = {{Candidate gene prioritization using graph embedding}},
  booktitle = {RIVF},
  year = {2020},
  doi = {10.1109/rivf48685.2020.9140776},
}

Loading related content...

Figure