Abstract
Drug combination therapies are critical in oncology to overcome resistance, yet tracking their prevalence in clinical registries is difficult due to a lack of explicit labeling. This study developed a natural language processing (NLP) scoring system to identify combination trials within ClinicalTrials.gov. Analyzing data from 2008 to 2013, the researchers found that while oncology has the highest rate of combination trials, the overall ratio of these trials stagnated or declined during the study period. Furthermore, NIH-funded studies were significantly more likely to explore drug combinations than industry-sponsored trials, highlighting a divergence in research priorities between public and private sectors.
Introduction
In the era of precision medicine, cancer is increasingly understood as a complex network of mutations rather than a single genetic error. This complexity often allows tumors to bypass single-target drugs, leading to treatment resistance. Consequently, combining multiple therapeutic agents to hit several pathways simultaneously has become a primary goal in oncology.
Despite its importance, there is a significant data gap: ClinicalTrials.gov, the world’s largest clinical trial registry, does not explicitly label trials as "combination" or "single agent." This information is often buried in free-text descriptions, hindering the ability of researchers to track trends or identify gaps in the therapeutic landscape. This study aims to bridge this gap by developing an automated tool to identify and categorize these trials.
Methodology and Data Processing
The researchers developed a rule-based NLP classifier to identify drug combination trials. While standard machine learning models like Random Forests were tested, a transparent, weighted keyword scoring system proved most effective for this specific dataset. The system assigns scores based on terms found in titles, summaries, and intervention descriptions (e.g., "combination" or "plus" add points, while "placebo" subtracts them).
To isolate oncology-specific trials, the team utilized Medical Subject Headings (MeSH) from the Aggregate Analysis of ClinicalTrials.gov (AACT) database. For trials lacking MeSH tags, they employed MetaMap to map free text to standardized Unified Medical Language System (UMLS) concepts. The final analysis focused on a subset of 8,233 combination trials and 14,652 oncology trials identified from an initial pool of over 170,000 records.
Key Trends in Oncology Combinations
The analysis revealed that oncology is a clear outlier in the use of combination therapy. Approximately 25.6% of oncology trials involve drug combinations, compared to much lower rates in cardiovascular disease (8.5%) or neurology (5.4%). However, the data also uncovered a "stagnation paradox": the ratio of combination trials to single-agent trials actually decreased from 29.5% in 2008 to 22.7% in 2012.
One significant finding was the disparity between public and private funding. NIH-funded trials are consistently more likely to explore drug combinations than industry-sponsored studies. This suggests that while academic and public researchers are focused on optimizing outcomes through multi-drug regimens, pharmaceutical companies may prioritize proprietary single-agent pipelines. The overall decline in combination ratios during the study period may also be linked to the 2010 FDA draft guidance on "codevelopment," which established stricter criteria for proving combination superiority, potentially increasing the perceived risk and cost for sponsors.
Results and Phase Distribution
The NLP system achieved high performance with a precision of 0.94 and a recall of 0.85. In terms of trial distribution, nearly half (45.5%) of oncology combination trials are in Phase 2, while only 1.5% reach Phase 4. This indicates a robust pipeline for testing efficacy, but a significant drop-off in combinations reaching post-market surveillance.
Discussion and Limitations
While providing a valuable roadmap for analyzing trial data, the study faced several limitations. Over 3,800 trials were excluded due to missing start dates, and the registry may underrepresent early-stage (Phase 0 and 1) trials where registration is not always mandatory. Future research will aim to stratify these trends by specific cancer types and identify the most frequently paired drug classes. By unlocking this hidden registry data, the authors hope to facilitate more rational and efficient clinical trial design for next-generation cancer therapies.
Summary updated