Introduction: Early identification of prostate cancer (PCa) patients that are most likely to develop a lethal disease is still a major hurdle. While for most patients, a radical treatment will result in the eradication of the disease, about 70% of patients with advanced PCa are diagnosed with bone metastasis, a lethal feature with limited treatment options. There is growing evidence suggesting that lipid metabolism cooperates with the genetic landscape of prostate cancer. Currently, several studies have shown that the metabolome of PCa could be used as non-invasive diagnostic and prognostic tool. In this study, we analyzed the plasma of 73 prostate cancer patients to identify metabolites with diagnostic or prognostic value. Methods: The plasma was collected during fasting and shortly before the patient received radical prostatectomy. The patients were then followed up for up to 10 years, to monitor the development of bone metastases. Targeted metabolomics was run on the pre-operative plasma. 10 different machine learning models were tested on the available metabolomics data-set. All patients provided informed consent for participation in this study. Results: By using machine learning algorithms, we identified a signature of 41 metabolites either positively or negatively associated with the development of bone metastatic PCa. Neural network and support vector machine yielded the best results for classification of metastasis based on the metabolites. Furthermore, linear support vector machine and Gaussian process generated the best regression models for determination of time-to-bone metastasis using absolute Spearman`s R correlation. Determination of both the classification of bone metastasis and time-to-metastasis were identified based on 41 metabolites in the Neural network and support vector machine algorithms. The top-four metabolites associated with bone metastatic progression were lysophosphatidylethanolamine(LPE 22:5/0:0)b, butyrylcarnitine (C4), pyruvic acid and dodecanoylcarnitine (C12), whereas monoacylglycerol (MAG 14:0/0:0), MAG (0:0/22:5), acetylornithine and MAG (0:0/14:0) were the bottom-four metabolites. Conclusions: In this study, we evaluated metabolites in the plasma of prostate cancer patients to determine if we could identify a predictive signature of metastatic progression. Our studies show that metabolic phenotypes have substantial clinical value. Given the context of precision medicine, we show the feasibility of a simple metabolomics test on the plasma of patients that could help stratify patient`s risk of lethal disease based on the presence/absence of signature metabolites. Further studies in a larger cohort will be required in order to test the signature metabolites found in our study to test the reliability for patient stratification. SOURCE OF Funding: Swiss Cancer Research Foundation (KFS-5510-02-2022-R)
