Abstract 5041: OncoTagger, a novel tool for accelerated and accurate literature-based curation of cancer related genes and variants

Abstract Cancer is a genetic disease, and a tremendous number of papers are published each year describing the role of individual genes and specific genomic variants in different cancer types, on cancer progression, and response to therapy. Databases such as OncoKB, COSMIC, and ClinVar effectively catalogue these variants. However, building such databases currently requires laborious manual curation, which does not scale to the speed at which new findings are published: curators need to first identify relevant papers to identify novel associations, and then read through the selected papers word-by-word to ascertain such associations. To support and accelerate this process, we developed OncoTagger, a first-of-its-kind natural language processing (NLP) powered tool that identifies cancer-related genes and variants from published abstracts and classifies them according to their oncogenicity. OncoTagger provides an abstract-level summary to identify important papers containing cancer genes and variants. Within each abstract, OncoTagger further provides a word(token)-level assessment, predicting words that are cancer genes or variants and their oncogenicity to facilitate curation. OncoTagger consists of two BioMedBert models finetuned and evaluated on OncoKB’s cancer biomarker summary data (3381 annotated abstracts with 22219 mentions of cancer genes, and 1291 abstracts with 3446 cancer variant mentions). One model performs the gene classification task: a token can be assigned to be (part of) the name of an oncogene, tumor suppressor gene, both, or none of the above; the other model carries out similar classifications for variants. To evaluate abstract summary accuracy, we determined that OncoTagger’s most frequently predicted class for each known cancer gene and variant token across the held-out test abstracts achieved 95.2% and 82.1% sensitivity, respectively. We then evaluated accuracy on the task of token classification. OncoTagger significantly outperformed LSTM and key-term matching models in identifying cancer genes and variants (overall accuracy 97.4%/99.4%, F1=72.3%/73.6%, 5.8pp/5.1pp better than the second-best model). OncoTagger achieved high sensitivity in detecting never-seen-before genes and variants (67.7%, 83.3% in oncogenes and tumor suppressor genes, respectively; 60.2% in oncogenic variants). Manual reviews of OncoTagger predictions are ongoing to further confirm OncoTagger’s ability to discover new cancer genes and variants currently not included in OncoKB. In summary, the NLP-powered OncoTagger tool accurately extracts cancer biomarkers. The scaled application of OncoTagger will help identify and curate cancer driver genes and variants from the literature, accelerating the translation of research into clinically implementable targets. Citation Format: Chenlian Fu, Pia Droop, Xing Wang, Hongxin Zhang, Nicole Fernandez, Calvin Lu, Abhilash Anand, Umair Ayub, Saahil Jain, Kritik Seth, Debyani Chakravarty, Ulf Leser, Quaid Morris, Nikolaus Schultz. OncoTagger, a novel tool for accelerated and accurate literature-based curation of cancer related genes and variants [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5041.