TL;DR
Researchers have developed an automated system to mine millions of scientific publications, revealing complex relationships between anti-cancer drugs and specific cancer types. This approach, validated against FDA approvals, experimental data, and patient survival, offers a powerful tool for drug repositioning and understanding cancer treatment landscapes.
What Problem Does This Paper Address?
The sheer volume of scientific literature on cancer and anti-cancer drugs makes it impossible for researchers and clinicians to manually extract all relevant information. This deluge of data hinders the identification of drug-cancer relationships, potential new uses for existing drugs (drug repositioning), and a comprehensive understanding of treatment strategies. The paper aims to bridge this gap by creating an automated system to systematically analyze this literature and uncover these crucial connections.
Methods
The study employs two primary text-mining strategies to analyze the relationship between 30 common cancer types and 270 anti-cancer drugs, using nearly 4 million PubMed abstracts:
- Classical Text Mining (Supervised): This approach uses named entity recognition (NER) to identify genes, proteins, and compounds. It then applies an over-representation analysis and Fisher's exact test to determine if specific genes/proteins or compound-cancer type associations are statistically significant within the literature. This method leverages prior knowledge about entities.
- Word Embeddings (Unsupervised): This technique transforms text into numerical vectors, allowing for the identification of similar words. By training word embeddings on the entire corpus, the study calculates cosine similarity between compound and cancer type vectors to infer relationships. This method is unsupervised, meaning it doesn't rely on pre-defined entity lists.
The consistency of the mined relations was validated using three independent methods: FDA drug approval data, experimentally determined IC-50 values (measuring drug sensitivity in cell lines), and clinical patient survival data.
Datasets
- Publications: Approximately 4 million PubMed abstracts related to cancer types and anti-cancer drugs were used.
- Cancer Types: Data for the 30 most frequent cancer types worldwide were sourced from the World Cancer Research Fund International.
- Anti-Cancer Drugs: A set of 270 compounds was compiled from the Genomics of Drug Sensitivity in Cancer (GDSC) database and the Cancer Cell Line Encyclopedia (CCEL).
- Validation Datasets:
- FDA Approval Data: A list of FDA-approved drugs for specific cancer types.
- IC-50 Values: Experimentally determined drug sensitivity data from the GDSC dataset.
- Patient Survival Data: Clinical datasets containing survival times and medication information, sourced from the Broad Institute.
Key Insights and Ideas
- Automated Discovery of Drug-Cancer Links: The study demonstrates that automated text mining can effectively identify and quantify the relationships between anti-cancer drugs and cancer types, mirroring findings from experimental and clinical data.
- Drug Repositioning Potential: The identification of "co-occurring compounds" within the literature provides a powerful avenue for suggesting new therapeutic uses for existing drugs. For instance, the analysis highlighted a strong literature connection between erlotinib and colon cancer, suggesting afatinib as a potential candidate for repositioning.
- General vs. Specific Drug Action: The analysis revealed that drugs can be broadly categorized into those targeting general cellular processes (like DNA replication) and those with more specific targets related to particular signaling pathways, which often correlates with their association with specific cancer types.
- Interactive Knowledge Base: The creation of a web-based knowledge base makes the complex findings accessible and explorable, allowing researchers to delve deeper into specific drug-cancer relationships and formulate new hypotheses.
Key Results
- High Agreement with Validation Data: The literature mining results showed strong agreement with FDA drug approvals (AUC of ~0.9 for prediction), experimentally determined IC-50 values (lower IC-50 values for highly associated drug-cancer pairs), and clinical patient survival data (drugs with higher literature support correlated with improved survival).
- Heatmap Visualization: A comprehensive heatmap visually represents the significance of drug-cancer relationships based on publication overlap, highlighting both broad associations (e.g., doxorubicin with many cancers) and highly specific ones (e.g., nilotinib with leukemia).
- Supervised vs. Unsupervised Performance: The supervised text mining approach (classical NER) generally performed better than the unsupervised word embedding approach in assessing drug-cancer relations, likely due to its direct incorporation of prior knowledge.
- Drug Combinations Identified: The analysis identified significant co-occurrences of drugs, suggesting potential synergistic combinations, such as 5-fluorouracil with irinotecan for colon cancer, and dabrafenib with trametinib.
Limitations and Future Work
- Abstracts vs. Full-Text: The study primarily used abstracts due to accessibility limitations. Full-text analysis could potentially yield more precise gene/protein extraction and improve word embedding performance.
- General Cancer Type Descriptions: The analysis used broad cancer type classifications (e.g., "breast cancer") rather than more specific subtypes (e.g., based on receptor status). Future work could incorporate this finer granularity if sufficient literature data becomes available.
- Data Availability: The study acknowledges that for some cancer types, generating survival curves was limited by the availability of patient data.
- Future Work: The authors plan to expand the knowledge base with more compounds and suggest that the identified co-occurring compounds can be used to propose drug repositioning strategies. The interactive knowledge base is intended for researchers to explore these findings and pursue their own research questions.
Summary updated