Systems and methods for predicting properties of peptides directly from mass spectrometry data

A method for predicting properties of peptides from mass spectrometry (MS) data of the peptides, the method comprising:

• i) receiving the MS data as input;
• ii) capturing sequence dependencies in the MS data;
• iii) generating an intermediate output using the sequence dependencies;
• iv) combining a feature vector of mass and charge values with the intermediate output to generate a combined feature vector;
• v) generating spectrum embeddings using the combined feature vector;
• vi) predicting the properties of the peptides based on the spectrum embeddings;
• vii) using the predicted properties of the peptides to diagnose a disease;
• viii) generating peptide embeddings from the peptides using two bidirectional long short-term memory networks;
• ix) estimating a confidence of the spectrum embeddings by ranking the peptide embeddings against the spectrum embeddings;
• x) determining a certainty of embedding location by introducing controlled augmentations to the MS data and measuring a variation in the spectrum embeddings;
• xi) measuring a training data density around spectrum embedding using a von Mises-Fisher Mixture Model;
• and
• xii) evaluating whether a first density of the MS data is consistent with a second density of the spectrum embeddings,
• the properties of the peptides being predicted directly from the MS data without any prior deduction or database search on the MS data.

A system for predicting properties of peptides from mass spectrometry (MS) data of the peptides, the system comprising:

• a processor;
• a display in operable communication with the processor;
• and
• a machine-readable medium in operable communication with the processor and the display and having instructions stored thereon that, when executed by the processor, perform the following steps:
• i) receiving the MS data as input;
• ii) capturing sequence dependencies in the MS data;
• iii) generating an intermediate output using the sequence dependencies;
• iv) combining a feature vector of mass and charge values with the intermediate output to generate a combined feature vector;
• v) generating spectrum embeddings using the combined feature vector;
• vi) predicting the properties of the peptides based on the spectrum embeddings;
• and
• vii) using the predicted properties of the peptides to diagnose a disease;
• viii) generating peptide embeddings from the peptides using two bidirectional long short-term memory networks;
• ix) estimating a confidence of the spectrum embeddings by ranking the peptide embeddings against the spectrum embeddings;
• x) determining a certainty of embedding location by introducing controlled augmentations to the MS data and measuring a variation in the spectrum embeddings;
• xi) measuring a training data density around spectrum embedding using a von Mises-Fisher Mixture Model;
• and
• xii) evaluating whether a first density of the MS data is consistent with a second density of the spectrum embeddings,
• the properties of the peptides being predicted directly from the MS data without any prior deduction or database search on the MS data,
• the MS data comprising mass to charge ratio (m/z), intensity, and index value,
• the properties of the peptides comprising length, missed cleavages, and modification status,
• the capturing of the sequence dependencies in the MS data comprising using two 8-headed self-attention layers,
• the predicting the properties of the peptides based on the spectrum embeddings comprising sending the spectrum embeddings through a first branch of two fully connected layers to predict a first property of the peptides, a second branch of two fully connected layers to predict a second property of the peptides, and a third branch of two fully connected layers to predict a third property of the peptides, and
• the generating an intermediate output using the sequence dependencies comprising sending the sequence dependencies through two fully connected layers to generate the intermediate output.