An evaluation methodology for machine learning-based tandem mass spectra similarity prediction

Untargeted tandem mass spectrometry serves as a scalable solution for the organization of small molecules. One of the most prevalent techniques for analyzing the acquired tandem mass spectrometry data (MS/MS) - called molecular networking - organizes and visualizes putatively structurally related compounds. However, a key bottleneck of this approach is the comparison of MS/MS spectra used to identify nearby structural neighbors. Machine learning (ML) approaches have emerged as a promising technique to predict structural similarity from MS/MS that may surpass the current state-of-the-art algorithmic methods. However, the comparison between these different ML methods remains a challenge because there is a lack of standardization to benchmark, evaluate, and compare MS/MS similarity methods, and there are no methods that address data leakage between training and test data in order to analyze model generalizability.

In this work, we present the creation of a new evaluation methodology using a train/test split that allows for the evaluation of machine learning models at varying degrees of structural similarity between training and test sets. We also introduce a training and evaluation framework that measures prediction accuracy on domain-inspired annotation and retrieval metrics designed to mirror real-world applications. We further show how two alternative training methods that leverage MS specific insights (e.g., similar instrumentation, collision energy, adduct) affect method performance and demonstrate the orthogonality of the proposed metrics. We especially highlight the role that collision energy plays in prediction errors. Finally, we release a continually updated version of our dataset online along with our data cleaning and splitting pipelines for community use.

It is our hope that this benchmark will serve as the basis of development for future machine learning approaches in MS/MS similarity and facilitate comparison between models. We anticipate that the introduced set of evaluation metrics allows for a better reflection of practical performance.