TL;DR
This paper introduces Pathomap, a novel computational approach that leverages word embeddings from biomedical literature to create a visual atlas of tissue-specific gene pathological roles. By analyzing millions of PubMed abstracts, Pathomap provides a comprehensive, tissue-level understanding of how genes contribute to disease, accelerating targeted investigations and potentially uncovering new disease-gene associations.
What Problem Does This Paper Address?
Understanding the complex interplay between genes and diseases is a monumental challenge. While we have made strides in mapping healthy human tissues at a cellular level, a similar comprehensive atlas for disease pathology is lacking. This gap exists due to genetic pleiotropy (one gene affecting multiple traits) and the heterogeneity of disease manifestations, which are often influenced by various clinical and demographic factors. Existing methods for disease-gene association are often manual, labor-intensive, and don't provide a clear tissue-specific perspective. The authors aim to bridge this gap by creating a computational tool that can systematically map the pathological roles of genes within specific tissues.
Methods
The core of Pathomap relies on Natural Language Processing (NLP) techniques, specifically word embeddings. The authors first curated a dataset of approximately 6.1 million PubMed abstracts that explicitly describe the pathological roles of genes in diseases. This focused dataset was then used to train a custom Word2vec model. Unlike generic word embeddings trained on broad biomedical literature, this disease-focused approach aims to capture more nuanced semantic relationships relevant to pathology. The trained embeddings allow for the calculation of a "Patho-score," which quantifies the association between a gene and a tissue. They also developed an abstract stratification strategy using BioSentVec and SVM models to filter relevant abstracts. For visualization and accessibility, they created a Shiny application and an R package.
Datasets
- PubMed Abstracts: Approximately 18 million abstracts were initially downloaded, from which ~6.1 million "patho-abstracts" (describing gene-disease associations) were selected for training.
- DisGeNET: Used as a benchmark for evaluating the performance of Pathomap's automated approach against manually curated disease-gene relationships.
- NCBI Human Genes List: Used for comparison of gene coverage between Pathomap and DisGeNET.
- GTEx Gene Expression Archives: Used for comparing gene expression levels with Patho-scores.
Key Insights and Ideas
- Context-Specific Embeddings Matter: Training word embeddings exclusively on disease-focused abstracts (patho-abstracts) proved more effective in capturing pathological relationships than using generic biomedical embeddings (like BioBERT). This highlights the importance of domain-specific learning for complex biological data.
- Gene Modules in Proximal Tissues: The analysis revealed distinct gene modules with reproducible pathological roles across anatomically proximal tissues. This suggests that certain biological processes and disease mechanisms might be shared among related organs.
- Uncovering Novel Associations: Pathomap successfully identified a potential association between the genes DNMT3A/BCOR and CYLD cutaneous syndrome (CCS) before this association was formally reported in the literature, demonstrating its power for knowledge synthesis and early discovery.
- Visualizing Pathological Landscape: The Pathomap tool provides an intuitive visual representation of a gene's pathological impact across different tissues, offering a high-level, empirical view that is both trustworthy and actionable.
Key Results
- Pathomap successfully extracted latent knowledge from ~6.1 million PubMed abstracts.
- The Patho-scores for experimentally validated disease-gene relationships from DisGeNET showed a significant right-skewed distribution compared to random associations, indicating high fidelity.
- Pathomap covers a substantial portion of genes and disease names, with about half of the DisGeNET annotated diseases being represented in the Pathomap embeddings.
- The heatmap analysis identified coherent gene modules with similar pathological roles across anatomically related tissues (e.g., brain regions, skeletal muscle and adipose tissue).
- The tool demonstrated its ability to predict domain-specific analogies, such as gene-disease relationships and organ-symptom associations.
Limitations and Future Work
- Species Segregation: The current study did not segregate articles by species, meaning the Patho-scores do not differentiate between human, mouse, or other primate data. The authors believe this is less of a nuisance as animal models are often used to mimic human diseases.
- No Specific Mechanism: Pathomap indicates the extent of a gene's involvement in disease within a tissue but does not provide specifics on the exact mechanism of involvement.
- Future Work: The authors plan to update the Shiny application and R package yearly to incorporate new discoveries. They suggest Pathomap can be used for diagnostic gene-panel selection, as an orthogonal approach for Gene Ontologies, and for prioritizing genes in large-scale loss-of-function studies.
Summary updated