Systems and methods for de novo peptide sequencing using deep learning and spectrum pairs

Inventors

Qiao, Rui • TRAN, Ngoc Hieu • XIN, Lei • Chen, Xin • SHAN, Baozhen • Ghodsi, Ali • Li, Ming

Assignees

BIOINFORMATICS SOLUTIONS Inc

Abstract

The present systems and methods are directed to de novo identification of peptide sequences from tandem mass spectrometry data. The systems and methods uses unconverted mass spectrometry data from which features are extracted. Using unconverted mass spectrometry data reduces the loss of information and provides more accurate sequencing of peptides. The systems and methods combine deep learning and neural networks to sequencing of peptides.

Core Innovation

The invention concerns a computer implemented system and corresponding method for de novo sequencing of peptides from a sample mass spectrometry data using neural networks. The system generates a probability measure for one or more candidates to a next amino acid in an amino acid sequence by using a plurality of layered nodes configured as an artificial neural network, and the neural network is trained on known mass spectrometry data representing theoretical fragment ion peaks of known sequences differing in length and differing by one or more amino acids.

The system receives sample mass spectrometry data comprising a plurality of coordinate data pairs representing m/z and intensity values of observed fragment ion peaks and receives second coordinate data representing theoretical fragment ion peaks. The layered nodes include at least one convolutional and/or fully connected layer for comparing at least one of the observed fragment ion peaks against at least one of the theoretical fragment ion peaks, and identifies the difference in m/z between the observed fragment ion peaks and the theoretical fragment ion peaks by constraining a difference tensor (D).

After the peptide-sequencing neural network produces output, the system obtains an input prefix representing a determined amino acid sequence of the peptide. The system identifies a next amino acid based on a candidate next amino acid having a greatest probability measure based on the output of the artificial neural network and the sample mass spectrometry data, and updates the determined amino acid sequence with the next amino acid. In the described architecture, an LSTM is used as a language model to iteratively predict the next amino acid symbol from an input prefix and update the peptide sequence, with learned representations including a difference tensor constraint and an m/z positional embedding.

Claims Coverage

The partial content provides two independent claims: one system claim and one method claim. Across these independent claims, the main inventive features involve neural-network-based comparison of observed and theoretical fragment ion peaks from sample mass spectrometry coordinate pairs, constrained difference-tensor matching of m/z differences, and iterative next-amino-acid probability prediction from an input prefix to update a determined amino acid sequence.

Together, the independent claims cover de novo peptide sequencing by using layered neural networks to compare observed and theoretical fragment ion peaks represented as m/z and intensity coordinate pairs, constrain a difference tensor (D) to identify matching theoretical fragment ion peaks, and iteratively select the next amino acid by choosing the candidate with the greatest probability measure based on an input prefix and the neural network outputs.

Stated Advantages

Documented Applications