Abstract Background: Identifying the primary site of origin for tumors in an accurate and timely manner is a critical step in treating cancer in an effective way. In cases of cancers of unknown primary, deep learning-based genomic classifiers that can improve diagnostic workflows and decrease inaccuracies have been validated for clinical use by previous studies, but reports of clinical utility are lacking. Use of molecular profiling for identification of site of origin has shown to be accurate in about 70-95% cases and can reveal actionable targets for treatment. We report our experience of consecutive cases where a commercial AI-based classification tool was used in routine practice at a high-volume community cancer center. Methods: Patient tumors, initially classified as cancers where the primary site was uncertain, were subjected to molecular profiling and AI-based genomic classification (GPSai) capable of recognizing a set of 90 distinct primary tumor sites and many histological subclasses. Molecular profiling consisted of whole exome and whole transcriptome sequencing of FFPE tumor samples, along with other predictive biomarker assays performed by a commercial laboratory (Caris Life Sciences). The identification of targetable mutations was reported, but variants of unknown significance were excluded when reporting genomic alterations. Results: We reviewed 320 samples from 313 patients between December 2022 and October 2025 who were profiled using GPSai, 77 (24%) of which were cancers with unknown primary. Those cases which were unable to be classified were further investigated by a molecular pathologist. Of the 243 (76%) samples that received a tumor type assignment by the classifier, non-small cell lung cancer was the most represented, with 41 of 243 (16.9%), followed by pancreatic adenocarcinoma, with 25 of 243 (10.3%), and colon cancer, with 21 of 243 (8.6%). The classifier also assigned classifications of miscellaneous rare cancer types (21 of 243, or 8.6%), including rhabdomyosarcoma, angiosarcoma, and thymic carcinoma. 161 of 313 (51.4%) patients were women, 152 of 313 (48.6%) were men, and 201 of 313 (64%) patients were 65 or older in age. Mutations were found in 173 of 320 (54%) samples, with the most common being in TP53, KRAS, and TERT. Further, 51 of 320 (16%) of samples analyzed had a high tumor mutational burden (TMB), 10 of 320 (3%) had a mismatch repair deficiency, and 45 of 320 (14%) samples were positive for PD-L1 (SP142) and 53 of 320 (16.6%) for PD-L1 (22c3). Conclusion: 243 of 330 (76%) patient samples were assigned a cancer diagnosis, which is consistent with the range reported in published studies to date (70-95%). Among those that could not be classified, targetable mutations were commonly found, including KRAS and EGFR. AI-based genomic classification of tumor site of origin is useful in clinical practice in identifying treatment options for patients. Citation Format: Carlos E. Zuazo, Michael J. Demeure, David R. Braxton, Sourat Darabi. Identifying the tumor site of origin using molecular profiling with an AI-based classifier abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 3875.
Zuazo et al. (Fri,) studied this question.