1. Machine Learning Predicts Renal Cell Carcinoma Status from Urine Using Multiplatform Metabolomics.
Link:ASMS_MAY2020_Poster_Bifarin_final
I presented a poster at the American Society for Mass Spectrometry Conference 2020 about using machine learning and urine-based metabolomics to detect kidney cancer. The study used machine learning techniques to predict kidney cancer status from metabolomic profiles generated through liquid chromatography/mass spectrometry and nuclear magnetic resonance. The cohort, comprising 82 kidney cancer patients and 174 healthy controls, was divided for training and validation of the machine learning model. The final machine learning model achieved a 98% accuracy rate in distinguishing between healthy controls and cancer patients. This research indicates that renal cell carcinoma diagnosis may be possible through routine urine tests in the future.
Link: ASMS_JUNE2022_Poster_Bifarin_final
My poster presentation at the 2022 American Society for Mass Spectrometry Conference revolved around using a Dicer-Pten Double-Knockout mouse model to study lipidome changes in high-grade serous carcinoma (HGSC), a lethal type of ovarian cancer. Ultra-high performance liquid chromatography-mass spectrometry (UHPLC-MS) was used to conduct serum lipidomic profiling. The study identified significant changes in lipid features over the disease progression stages. Machine learning was applied for classifying different stages of ovarian cancer progression, with the highest classification performance at 46-60% of the mouse's lifetime. This longitudinal study provided a first-of-its-kind insight into lipidome dynamics in HGSC, suggesting significant perturbations in cell membrane stability, cellular proliferation, and survival.
Link: Bifarin_ASMS_2023_Poster_final
The selection of optimal machine learning (ML) models for MS-based metabolomics is crucial but often involves tedious evaluation. Automated Machine Learning (AutoML) can automate this process, but the outputs can be difficult to understand, necessitating the need for complex model interpretation. AutoSklearn was used for AutoML model selection, models were interpreted using the KernelSHAP method, and the pipeline was tested on a renal cell carcinoma (RCC) urine-based metabolomics LC-MS dataset.