Reproducibility and Environmental Efficiency of Metabolomics Cancer Modeling

Sustainability in the context of machine learning (ML) plays an important role for accessible models by both researchers as well as clinicians. This article describes a reproducibility study on PiDeeL, a metabolic-pathway-informed deep learning model. It serves to test the hypothesis that the requirement of a simple provision of all digital artifacts is not sufficient to reproduce the computational experiment(s). The reproduction and modification of the computational model foundational to the previous findings shall promote documentation and evaluation of existing scientific models and confirm their applicability. The modification of the original model is based on measuring emissions of training machine learning models using CodeCarbon. Two different systems with different CPU as well as GPU specifications and Windows Subsystem Linux could be tested after guide and code adaptions due to initial incomplete replication attempts given the threshold of computation completion without error message(s). Emissions equivalent to 0.3–0.6 kg of CO2 per run were shown. Encountered issues along the replication attempts call for refined guidelines on documentation and processing of computational approaches in scientific studies by publishers as well as the scientific community. Thorough peer review including algorithmic reproduction would be necessary to ensure model reusability.