Enhancing constituent estimation in nucleic acid mixture models using spectral annealing inference and MS/MS information

Mass spectrometry (MS) is a powerful analytical technique employed for a variety of applications including drug development, quality assurance, food inspection, and monitoring environmental pollutants. Recently, in the production of actively developed antibody and nucleic acid pharmaceuticals, impurities with various modifications have been generated. These impurities can lead to a decrease in drug stability, pharmacokinetics, and efficacy, making it crucial to distinguish between them. We previously modeled mass spectrometry for each possible number of constituents in a sample, using parameters such as monoisotopic mass and ion counts, and employed stochastic variational inference to determine the optimal parameters and the maximum posterior probability for each model. By comparing the maximum posterior probabilities among models, we selected the optimal number of constituents and inferred their corresponding monoisotopic masses and ion counts. However, MS spectra are sparse and predominantly flat, which can lead to vanishing gradients when using simple optimization techniques. To solve this problem, using MCMC as in our previous studies would take a very long time. To address this difficulty, in this study, we blur the comparative spectra and gradually reduce the blur to prevent vanishing gradients while inferring accurate values. Furthermore, we incorporate MS/MS spectra into the model to increase the amount of information available for inference, thereby improving the accuracy of parameter inference. This modification improved the mass error from an average of 1.348 Da–0.282 Da. Moreover, the required time, even including the processing of additional five MS/MS spectra, was reduced to less than half.