TL;DR
This paper introduces PhenotypeXpression (PhenoX), a novel computational tool that leverages public gene expression data and scientific literature to rapidly identify and characterize subtypes of diseases. By integrating these two data sources, PhenoX aims to provide more granular disease classifications, which is crucial for advancing personalized medicine and tailoring treatments.
What Problem Does This Paper Address?
The dream of personalized medicine hinges on accurately classifying and subtyping diseases. However, current disease classifications often lack the granularity needed to capture the subtle differences in how diseases manifest at a molecular level, particularly in terms of gene expression. This makes it challenging for clinicians to develop highly tailored treatment strategies. While vast amounts of gene expression data and biomedical literature are publicly available, integrating them to discover and define these finer disease subtypes has been a significant hurdle. PhenoX seeks to bridge this gap by providing a systematic way to mine these resources for disease subclassification.
Methods
PhenoX employs a multi-pronged approach:
- Gene Expression Data Aggregation and Clustering: The tool first queries the Gene Expression Omnibus (GEO) database for relevant gene expression datasets (GDS) associated with a given disease term. These GDS are then filtered and hierarchically clustered using an approximately unbiased method with bootstrap probabilities. The clustering is based on the presence or absence of differentially expressed genes, aiming to identify distinct groups of GDS that represent potential disease subtypes.
- Literature Mining for Phenotypic Profiling: For each identified cluster of GDS, PhenoX retrieves associated PubMed articles. It then uses Natural Language Processing (NLP) techniques, specifically Named Entity Recognition (NER) with the Spacy Lookup extension, to extract Human Phenotype Ontology (HPO) and Disease Ontology (DOID) terms from the titles, abstracts, and keywords of these articles.
- Term Co-occurrence Analysis: The frequency of these HPO and DOID terms within each cluster's literature corpus is analyzed. This allows PhenoX to generate a "phenotypic profile" for each subtype, characterized by the co-occurrence of specific terms, and to visualize these differences using word clouds.
Datasets
PhenoX utilizes the following public datasets:
- Gene Expression Omnibus (GEO): A repository of high-throughput gene expression data.
- PubMed: A database of biomedical literature.
- MeSH (Medical Subject Headings): Used for disease term matching and hierarchy navigation.
- Human Phenotype Ontology (HPO): A standardized vocabulary for describing phenotypic abnormalities.
- Disease Ontology (DOID): A comprehensive ontology for human diseases.
The paper demonstrates PhenoX's utility on the disease "Psoriasis," identifying two distinct clusters.
Key Insights and Ideas
- Synergistic Power of Gene Expression and Literature: The core insight is that combining gene expression data with textual information from scientific literature provides a more comprehensive understanding of disease subtypes than either source alone. Gene expression reveals molecular differences, while literature provides clinical and phenotypic context.
- Automated Subtype Discovery: PhenoX automates a complex process that would otherwise be manual and time-consuming, enabling rapid exploration of disease heterogeneity.
- Phenotypic Fingerprinting: The generation of phenotypic profiles through literature mining offers a novel way to characterize and differentiate disease subtypes, moving beyond purely molecular definitions.
- Potential for Uncovering Comorbidities/Confounders: The identified subtypes, particularly in the Psoriasis example, suggest potential links to other conditions, hinting at underlying shared mechanisms or confounding factors.
Key Results
The paper demonstrates PhenoX's capability by applying it to Psoriasis, successfully identifying two distinct clusters of gene expression data.
- Psoriasis Subtypes: Cluster 1, in particular, was found to contain a number of epidermal inflammatory or infectious diseases, including oral candidiasis, scabies, and mucositis, suggesting a potential subtype characterized by these conditions.
- Visual Outputs: PhenoX generates several useful visualizations, including hierarchical clustering dendrograms, differential gene expression heatmaps, pairwise distance graphs, and word clouds of associated terminology.
- Computational Performance: The tool runs efficiently, with a typical execution time of 1-2 hours on a personal laptop for complex MeSH terms like Psoriasis and Sepsis, with clustering taking less than ten minutes.
Limitations and Future Work
- GEO Data Limitations: A significant limitation acknowledged is the availability and curation of data in GEO. The "up down genes" filter, used for initial data selection, is described as "ad hoc" and not an absolute determination of differential expression. A more thorough analysis of each GDS is beyond the scope of this project.
- Data Availability: The availability of representative GDS (at least two required) can be a limiting factor; some disease terms may not yield sufficient data for clustering.
- Future Directions:
- Exploratory Hypotheses: The identified subtypes are presented as exploratory, inviting further research into using subtype-associated genes for cohort selection, exploring treatment options, and identifying potential comorbidities.
- Interface Improvements: Enhancements to PhenoX's interface could allow for more flexible data exploration and parallel processing, particularly for NER, to further shorten runtime.
- Web-Hosted Interface: The tool's efficiency and small output files make it ideal for a web-hosted graphical interface for researchers without extensive computational backgrounds.
- RNA-seq Integration: PhenoX can be modified to accept gene expression data from newer technologies like RNA-seq, which are generally easier to analyze programmatically and correct for batch effects.
- Deeper Phenotype Analysis: Future work could involve investigating deep phenotypes distinct within the identified subgroups and exploring eQTLs (expression quantitative trait loci) in these subpopulations.
Summary updated