Cross-species transfer learning for liquid biopsy-based cancer detection

About This Grant

PROJECT SUMMARY There is a crucial need in the field of advanced diagnostic tool development for early detection of rare cancers such as osteosarcoma (OSA). Specifically, there is a need to develop platforms that can overcome the limitations of small sample sizes inherent in rare cancers while maintaining generalizability across patient populations for high test accuracy to improve patient outcomes. The long-term objective of this project is to develop a Raman spectroscopy-based liquid biopsy platform which uses a model trained on readily available companion animal biological samples as a basis for cross-species validation using smaller human data sets. To accomplish this objective, the research team will undertake two specific aims. First, Raman spectral signatures of canine OSA will be defined and validated as a foundation for cross-species diagnostic modeling. This will be done by using self-supervised models to train on Raman data generated from 100 canine plasma samples and subsequently using supervised classifiers to detect OSA. Biological interpretability will be validated via targeted lipidomic and metabolomic profiling. Second, the research team will develop and validate a cross-species transfer learning framework by performing Raman spectroscopy on more than 1,000 unannotated canine plasma samples and fine-tuning models on rare cancer–specific data in human patients. By accomplishing these aims, a comparative transfer learning framework will be established that enables label- efficient diagnostic development for rare cancers using abundant veterinary data. This platform has potential for broad extension to other rare cancers with canine analogs, providing a scalable path to early, noninvasive cancer detection in settings where human samples are limited.