Methods and systems for de novo peptide sequencing using deep learning

Inventors

SHAN, Baozhen • TRAN, Ngoc Hieu • Li, Ming • XIN, Lei • Zhang, Xianglilan

Assignees

BIOINFORMATICS SOLUTIONS Inc

Abstract

The present systems and methods introduce deep learning to de novo peptide sequencing from tandem mass spectrometry data. The systems and methods achieve improvements in sequencing accuracy over existing systems and methods and enables complete assembly of novel protein sequences without assisting databases. The present systems and methods are re-trainable to adapt to new sources of data and provides a complete end-to-end training and prediction solution, which is advantageous given the growing massive amount of data. The systems and methods combine deep learning and dynamic programming to solve optimization problems.

Core Innovation

The invention provides a computer implemented system for de novo sequencing of peptides from mass spectrometry data using neural networks. A processor and at least one memory provide a plurality of layered nodes configured to form an artificial neural network that receives a mass spectrometry spectrum data input and generates a probability measure for one or more candidates to a next amino acid in an amino acid sequence.

The artificial neural network is trained on known mass spectrometry spectrum data containing a plurality of known fragment ion peaks of known sequences that differ in length and differ by one or more amino acids. At least one convolutional layer filters mass spectrometry spectrum data to detect fragment ion peaks corresponding to amino acid sequences of various lengths, and at least one fully-connected layer identifies pairs of fragment ion peaks by fitting the plurality of known fragment ion peaks against the mass spectrometry spectrum data.

In additional architectures, the layered nodes include a first set and a second set of layered nodes. The first set comprises at least one convolutional layer and at least one fully-connected layer for fitting fragment ion peaks and identifying fragment-ion peak relationships, and the second set comprises one or more layers comprising a convolutional neural network (CNN) that identifies the presence of amino acids and one or more layers comprising a recurrent neural network (RNN) that predicts the next amino acid by vector embedding the output vectors.

The system obtains an input prefix representing a determined amino acid sequence, identifies a next amino acid from a candidate next amino acid having a greatest probability measure based on the output of the artificial neural network and the mass spectrometry spectrum data, and updates the determined amino acid sequence with the next amino acid. The described approach supports bidirectional sequencing and uses global/constraint-based optimization and decoding based on candidate sequences, including beam searching and mass-suffix filtering.

Claims Coverage

The independent claims cover a computer implemented system and a method for de novo peptide sequencing from mass spectrometry data using neural networks that output probability measures for candidate next amino acids, with CNN-based fragment-ion peak filtering, fully-connected pairing logic, optional CNN/RNN embedding pathways, and iterative prefix extension based on greatest probability. The independent claims specify two CNN-based pathways that can share a common last fully-connected layer to output the probability measure.

Across the independent claims, the core coverage is an artificial neural network system and method that filters mass spectrometry spectrum data with convolutional layers to detect fragment ion peaks, uses fully-connected layers to identify fragment-ion peak relationships offset by ±1 amino acid, and outputs a probability measure for candidate next amino acids used to iteratively extend an input prefix. Claim coverage further includes an integrated architecture with a CNN/RNN pathway for amino-acid presence vector generation and next-amino-acid prediction, combined with the fragment-ion pathway through shared last fully-connected layers to output the candidate next-amino-acid probability measures.

Stated Advantages

Documented Applications