Pideel: metabolic pathway-informed deep learning model for survival analysis and pathological classification of gliomas

Abstract

The real-time assessment of tumor characteristics during surgery is crucial, offering the potential to create a feedback mechanism for surgeons. Such feedback enables surgeons to make more informed decisions regarding the extent of tumor resection and whether to adopt a more aggressive or conservative approach. Al-though metabolomics-based tumor pathology prediction methods exist, their predictive performance is often constrained by the limited size of available datasets. Additionally, the feedback about the tumor tissue could be enhanced in terms of content and accuracy. This thesis introduces PiDeeL, a metabolic pathway-informed deep learning model designed to perform survival analysis and pathology assessment based on metabolite concentrations. We demonstrate that integrating pathway information into the model architecture significantly reduces parameter complexity while enhancing survival analysis and pathological classification per-formance. Our results indicate that PiDeeL improves tumor pathology prediction performance over state-of-the-art methods, achieving a 3.38% improvement in the Area Under the ROC Curve and a 4.06% improvement in the Area Under the Precision-Recall Curve. Similarly, regarding the time-dependent concordance index (c-index), PiDeeL exhibits superior survival analysis performance, with a 4.3% improvement compared to current leading methods. Furthermore, importance analyses conducted on input metabolite features and pathway-specific neurons of PiDeeL provide valuable insights into tumor metabolism. Applying this model in the surgical setting will assist surgeons in dynamically adjusting their surgical plans, ultimately leading to more accurate prognosis estimates tailored to the specifics of the surgical procedure.