Abstract
PURPOSE: Cell surface-targeted therapies (CSTs) are a rapidly expanding class of cancer treatments with high specificity and reduced toxicity. Matching patients who express specific targets to CST clinical trials remains challenging because of complex eligibility criteria, diverse targets, and the absence of centralized, up-to-date trial databases. These gaps limit patient access and contribute to poor trial accrual. METHODS: We developed a large language model (LLM)-driven pipeline to identify and annotate CST clinical trials. Using a two-pronged approach, LLMs extracted target information from ClinicalTrials.gov and the National Cancer Institute Drug Database. Eight LLMs, including GPT-4o and several open-source models, were benchmarked against manually curated data sets of 814 CST trials and 814 non-CST trials. We evaluated model performance at target and trial levels and analyzed sources of error. We also provide an up-to-date database of open CST trials and their targets from the >100,000 total oncology clinical trials in ClinicalTrials.gov. RESULTS: GPT-4o achieved the highest accuracy in identifying CST trials (96.5%) and their targets (89.5%). Combining data sources improved performance, and accuracy increased with later trial phases. Most errors stemmed from vague therapy descriptions or string-matching issues. The model matched 94% of US trials and >95% of trials globally, with exceptions in China and New Zealand. In predicting cell surface localization, Gemma 3:27b and MedLlama3 correctly labeled all known clinical cell surface targets although performance varied beyond the most well-known CSTs. CONCLUSION: Our LLM-based approach enables real-time, automated matching of patients to CST clinical trials, addressing major barriers to enrollment and expanding trial accessibility. Errors were uncommon, and performance is poised to improve as LLMs evolve. Optimizing patient-trial matching for CSTs can improve both patient benefit and trial success.