MassSpecGym: A benchmark for the discovery and identification of molecules • MassSpecGym introduces a comprehensive benchmark and dataset for molecular discovery from mass spectrometry (MS/MS) data, addressing limitations of existing datasets in scope, standardization, and data leakage prevention. • With 231,000 high-quality labeled MS/MS spectra representing 29,000 unique molecules, it is the largest open-access resource for MS/MS data, including three core challenges: de novo molecular structure generation, molecule retrieval, and spectrum simulation. • The benchmark provides innovative splitting techniques to prevent data leakage and evaluates model performance in a generalization-demanding environment, setting a new standard for machine learning tasks in mass spectrometry. • MassSpecGym offers three tasks to support different research goals: predicting molecular structures from MS/MS spectra, retrieving known molecules from spectral libraries, and simulating spectra for any given molecule. • This tool is accessible through a user-friendly interface built on PyTorch Lightning and Hugging Face, promoting reproducibility and simplifying model evaluation and benchmarking within the broader machine learning community. @tomas_pluskal @jjjvanderhooft @me_datapoint @sohahassoun @bittremieux @WishartLab @BoWang87 @rschmid1789 @corinnabrungs @boeckerlab @roman_bushuiev @AntonBushuiev 💻Code: github.com/pluskal-lab/MassS… 📜Paper: arxiv.org/abs/2410.23326 #MassSpectrometry #MachineLearning #DrugDiscovery #Metabolomics #Benchmark