Molecular Complexity to Biosignatures: A Machine Learning Pipeline that Connects Mass Spectrometry to Molecular Synthesis and Reaction Networks

Abstract

The search for life beyond Earth is complicated by the lack of a consensus on what life is – especially when considering potential forms of life not resembling anything known on Earth. Agnostic means of assessing samples for evidence of life are needed to address this challenge. Information encoded within the atoms and bonds of a molecule can be used to generate agnostic metrics of complexity. The distributions of complexity metrics for chemical mixtures involving biological processes have been hypothesized to be different from those produced by abiotic or prebiotic chemical reactions (Marshall et al. 2021). Complexity metrics, rooted in Shannon Entropy (Bertz 1981; Böttcher 2016) and Assembly Theory (Marshall et al., 2017, 2021), rely on knowledge of the precise structures of molecules and time-consuming human-expert-based analysis decoupled from real-time instrumental measurements onboard robotic missions. In addition, leveraging these metrics requires intensive – often intractable – computations that are infeasible for real-time, on-probe investigations. We propose light-weight, flexible neural network models, trainable from publicly available datasets that can be employed to predict molecular structures and their complexity metrics from mass spectra. We show that with careful selection of datasets, the ML-based approach can learn characteristics of experimental data and digital representation of molecules. This enables rapid, accurate prediction of molecular complexity from mass spectra. Such data pipelines may open new doors for critical robotic missions where autonomous decision-making is required, empowering rapid biosignature screening tasks and in situ fingerprinting of prebiotic molecular reaction networks.