Artificial neural network models for prediction of de novo sequencing of chains of amino acids
The present invention relates to proteomics, and techniques for predicting de novo sequencing of chains of amino acids, such as peptides, proteins, or combinations thereof. Particularly, aspects of the present invention are directed to a computer implemented method that includes obtaining a digital representation of a mass spectrum, the digital representation including a plurality of container elements, encoding, using an encoder portion of a bidirectional recurrent neural network of long short term memory cells and gated recurrent unit cells, each container element as an encoded vector, decoding, using a decoder portion of the bidirectional recurrent neural network, each of the encoded vectors into a sequence of amino acids; and recording the sequence of amino acids as a multi-dimensional data set of amino acids types and a probability of each of the amino acid types in each position of the complete amino acid sequence.
1. A method for identifying an unknown peptide or protein in a sample, the method comprising:
obtaining the sample comprising a mixture of proteins and molecules;
preprocessing the mixture of proteins and molecules to isolate the unknown peptide or protein from the mixture of proteins and molecules;
analyzing, using a mass spectrometer, the isolated peptide or protein to obtain a mass spectrum comprising a two-dimensional data set of mass and intensity values for each fragment ion from the isolated peptide or protein, wherein the analyzing comprises determining the mass and intensity values for each fragment ion from a m/z ratio and abundance detected by the mass spectrometer;
generating, by a computing device, a digital representation of the mass spectrum, the digital representation including a plurality of container elements, and each container element of the plurality of container elements is an object that stores other objects including a one-hot vector, a set of m/z values and respective abundancies, or a combination thereof that uniquely identifies each ion fragment within the mass spectrum;
inputting, by the computing device, the digital representation of the mass spectrum into an encoder-decoder network comprising an encoder portion and a decoder portion, wherein:
the encoder portion comprises a first bidirectional recurrent neural network of a first set of long short term memory cells and gated recurrent unit cells that are trained to map a variable-length source sequence defined by each container element to a fixed-dimensional vector representation,
the first bidirectional recurrent neural network is configured to process the variable-length source sequence from start to end and then process the variable-length source sequence from end to start,
the decoder portion comprises a second bidirectional recurrent neural network of a second set of long short term memory cells and gated recurrent unit cells that are trained to map the fixed-dimensional vector representation back to a variable-length amino acid sequence,
the second bidirectional recurrent neural network is configured to process the fixed-dimensional vector representation from start to end and then process the fixed-dimensional vector representation from end to start, and
the encoder-decoder network is initially trained on a first set of training data comprising standard amino acid residues, and then fine-tuned on a second set of training data comprising post-translational modifications of peptide or protein chains;
encoding, by the computing device and using the encoder portion, each container element as the fixed-dimensional vector representation, which is an m-dimensional or one-dimensional vector of m elements, wherein m corresponds to a number of different m/z values and abundances for n-most abundant ions;
appending, by the computing device, a set of metadata features to each container element or the fixed-dimensional vector representation for each container element, wherein the set of metadata features include: ionization method, mass spectrometer type, fragmentation method, fragmentation energy, peptide charge state, peptide mass-over-charge ratio, peptide's retention time, or a combination thereof;
generating, by the computing device using an attention model, a context vector for each of the fixed-dimensional vector representations based on intermediate outputs from the encoder portion from each step of encoding each of the fixed-dimensional vector representations, wherein the context vector is a weighted arithmetic mean of argument values, and weights are chosen according to relevance of each argument value given a context;
decoding, by the computing device and using the decoder portion, each of the fixed-dimensional vector representation into a variable-length amino acid sequence based on the set of metadata features and the context vector, wherein the variable-length amino acid sequence is represented as a multi-dimensional data set of amino acids types and probability of each amino acid type in each position of a sequence; and
identifying, by the computing device and based on the variable-length amino acid sequence, the unknown peptide or protein in order to characterize aspects including a complete sequence, structure, and possible function of the unknown peptide or protein, wherein the identifying comprises:
combining the variable-length amino acid sequences into a complete sequence of amino acids representing the unknown peptide or protein,
comparing the complete sequence of amino acids to a reference proteome, and
identifying a known peptide or protein based on the comparison.
2. The method of claim 1 , further comprising:
injecting the isolated peptide or protein into the mass spectrometer;
ionizing the injected peptide or protein in a gas phase using electrospray ionization (ESI) or matrix-assisted laser desorption/ionization (MALDI); and
accelerating the ionized peptides or proteins in the gas phase are accelerated in an electric or magnetic field for analysis,
wherein the preprocessing comprises using gel electrophoresis and/or liquid chromatography to isolate the unknown peptide or protein from the mixture of proteins and molecules, and wherein the fixed-dimensional vector representation is an m-dimensional or one-dimensional vector of m elements, where m corresponds to a number of different m/z values and abundances for the n-most abundant ions, and wherein the decoding comprises mapping the m-dimensional or one-dimensional vector of m elements back to the sequence of amino acids while taking into consideration a set of metadata features and a context vector.
3. The method of claim 1 , wherein the identifying further comprises:
matching one or more portions of the complete sequence of amino acids to one or more corresponding portions of a known peptide sequence;
determining a probability that the match between the one or more portions of the complete sequence of amino acids and the one or more corresponding portions of the known peptide sequence is indicative of a presence of the known peptide sequence in the sample; and
when the probability is greater than a predetermined threshold, obtaining an identifier of the corresponding peptide or protein sequence from the reference proteome,
wherein the unknown peptide or protein is identified based on the identifier.
4. The method of claim 1 , wherein each long short term memory cell comprises three gates and a cell unit, wherein each of the three gates is configured to use a sigmoid activation, and wherein a cell state is determined by transforming data associated with at least one gate using a hyperbolic tangent function.
5. The method of claim 1 , wherein each of the first and the second bidirectional recurrent neural network comprises at least two sets of hidden layers, wherein each set of the hidden layers is configured to processing data in a different direction.
6. A non-transitory machine readable storage medium having instructions stored thereon that when executed by one or more processors cause the one or more processors to perform a method comprising:
obtaining a sample comprising a mixture of proteins and molecules;
preprocessing the mixture of proteins and molecules to isolate an unknown peptide or protein from the mixture of proteins and molecules;
analyzing, using a mass spectrometer, the isolated peptide or protein to obtain a mass spectrum comprising a two-dimensional data set of mass and intensity values for each fragment ion from the isolated peptide or protein, wherein the analyzing comprises determining the mass and intensity values for each fragment ion from a m/z ratio and abundance detected by the mass spectrometer;
generating a digital representation of the mass spectrum, the digital representation including a plurality of container elements, wherein each container element of the plurality of container elements is an object that stores other objects including a one-hot vector, a set of m/z values and respective abundancies, or a combination thereof that uniquely identifies each ion fragment within the mass spectrum;
inputting the digital representation of the mass spectrum into an encoder-decoder network comprising an encoder portion and a decoder portion, wherein:
the encoder portion comprises a first bidirectional recurrent neural network of a first set of long short term memory cells and gated recurrent unit cells that are trained to map a variable-length source sequence defined by each container element to a fixed-dimensional vector representation,
the first bidirectional recurrent neural network is configured to process the variable-length source sequence from start to end and then process the variable-length source sequence from end to start,
the decoder portion comprises a second bidirectional recurrent neural network of a second set of long short term memory cells and gated recurrent unit cells that are trained to map the fixed-dimensional vector representation back to a variable-length amino acid sequence,
the second bidirectional recurrent neural network is configured to process the fixed-dimensional vector representation from start to end and then process the fixed-dimensional vector representation from end to start, and
the encoder-decoder network is initially trained on a first set of training data comprising standard amino acid residues, and then fine-tuned on a second set of training data comprising post-translational modifications of peptide or protein chains;
encoding, using the encoder portion, each container element as the fixed-dimensional vector representation, which is an m-dimensional or one-dimensional vector of m elements, wherein m corresponds to a number of different m/z values and abundances for n-most abundant ions;
appending a set of metadata features to each container element or the fixed-dimensional vector representation for each container element, wherein the set of metadata features include: ionization method, mass spectrometer type, fragmentation method, fragmentation energy, peptide charge state, peptide mass-over-charge ratio, peptide's retention time, or a combination thereof;
generating, using an attention model, a context vector for each of the fixed-dimensional vector representations based on intermediate outputs from the encoder portion from each step of encoding each of the fixed-dimensional vector representations, wherein the context vector is a weighted arithmetic mean of argument values, and weights are chosen according to relevance of each argument value given a context;
decoding, using the decoder portion, each of the fixed-dimensional vector representation into a variable-length amino acid sequence based on the set of metadata features and the context vector, wherein the variable-length amino acid sequence is represented as a multi-dimensional data set of amino acids types and probability of each amino acid type in each position of a sequence; and
identifying, based on the variable-length amino acid sequence, the unknown peptide or protein in order to characterize aspects including a complete sequence, structure, and possible function of the unknown peptide or protein, wherein the identifying comprises:
combining the variable-length amino acid sequences into a complete sequence of amino acids representing the unknown peptide or protein,
comparing the complete sequence of amino acids to a reference proteome, and
identifying a known peptide or protein based on the comparison.
7. The non-transitory machine readable storage medium of claim 6 , wherein the preprocessing comprises using gel electrophoresis and/or liquid chromatography to isolate the unknown peptide or protein from the mixture of proteins and molecules, and wherein the fixed-dimensional vector representation is an m-dimensional or one-dimensional vector of m elements, where m corresponds to a number of different m/z values and abundances for the n-most abundant ions, and wherein the decoding comprises mapping the m-dimensional or one-dimensional vector of m elements back to the sequence of amino acids while taking into consideration a set of metadata features and a context vector.
8. The non-transitory machine readable storage medium of claim 6 , wherein the identifying further comprises:
matching one or more portions of the complete sequence of amino acids to one or more corresponding portions of a known peptide sequence;
determining a probability that the match between the one or more portions of the complete sequence of amino acids and the one or more corresponding portions of the known peptide sequence is indicative of a presence of the known peptide sequence in the sample; and
when the probability is greater than a predetermined threshold, obtaining an identifier of the corresponding peptide or protein sequence from the reference proteome,
wherein the unknown peptide or protein is identified based on the identifier.
9. The non-transitory machine readable storage medium of claim 6 , wherein each long short term memory cell comprises three gates and a cell unit, wherein each of the three gates is configured to use a sigmoid activation, and wherein a cell state is determined by transforming data associated with at least one gate using a hyperbolic tangent function.
10. The non-transitory machine readable storage medium of claim 6 , wherein each of the first and the second bidirectional recurrent neural network comprises at least two sets of hidden layers, wherein each set of the hidden layers is configured to processing data in a different direction.
11. A system comprising:
one or more processors; and
a non-transitory machine readable storage medium containing instructions which, when executed on the one or more data processors, cause the one or more data processors to perform actions including:
instructing to obtain a sample comprising a mixture of proteins and molecules;
instructing to preprocess the mixture of proteins and molecules to isolate an unknown peptide or protein from the mixture of proteins and molecules;
instructing to analyze, using a mass spectrometer, the isolated peptide or protein to obtain a mass spectrum comprising a two-dimensional data set of mass and intensity values for each fragment ion from the isolated peptide or protein, wherein the analyzing comprises determining the mass and intensity values for each fragment ion from a m/z ratio and abundance detected by the mass spectrometer;
generating a digital representation of the mass spectrum, the digital representation including a plurality of container elements, and each container element of the plurality of container elements is an object that stores other objects including a one-hot vector, a set of m/z values and respective abundancies, or a combination thereof that uniquely identifies each ion fragment within the mass spectrum;
inputting the digital representation of the mass spectrum into an encoder-decoder network comprising an encoder portion and a decoder portion, wherein:
the encoder portion comprises a first bidirectional recurrent neural network of a first set of long short term memory cells and gated recurrent unit cells that are trained to map a variable-length source sequence defined by each container element to a fixed-dimensional vector representation,
the first bidirectional recurrent neural network is configured to process the variable-length source sequence from start to end and then process the variable-length source sequence from end to start,
the decoder portion comprises a second bidirectional recurrent neural network of a second set of long short term memory cells and gated recurrent unit cells that are trained to map the fixed-dimensional vector representation back to a variable-length amino acid sequence,
the second bidirectional recurrent neural network is configured to process the fixed-dimensional vector representation from start to end and then process the fixed-dimensional vector representation from end to start, and
the encoder-decoder network is initially trained on a first set of training data comprising standard amino acid residues, and then fine-tuned on a second set of training data comprising post-translational modifications of peptide or protein chains;
encoding, using the encoder portion, each container element as the fixed-dimensional vector representation, which is an m-dimensional or one-dimensional vector of m elements, wherein m corresponds to a number of different m/z values and abundances for n-most abundant ions;
appending a set of metadata features to each container element or the fixed-dimensional vector representation for each container element, wherein the set of metadata features include: ionization method, mass spectrometer type, fragmentation method, fragmentation energy, peptide charge state, peptide mass-over-charge ratio, peptide's retention time, or a combination thereof;
generating, using an attention model, a context vector for each of the fixed-dimensional vector representations based on intermediate outputs from the encoder portion from each step of encoding each of the fixed-dimensional vector representations, wherein the context vector is a weighted arithmetic mean of argument values, and weights are chosen according to relevance of each argument value given a context;
decoding, using the decoder portion, each of the fixed-dimensional vector representation into a variable-length amino acid sequence based on the set of metadata features and the context vector, wherein the variable-length amino acid sequence is represented as a multi-dimensional data set of amino acids types and probability of each amino acid type in each position of a sequence; and
identifying, based on the variable-length amino acid sequence, the unknown peptide or protein in order to characterize aspects including a complete sequence, structure, and possible function of the unknown peptide or protein, wherein the identifying comprises:
combining the variable-length amino acid sequences into a complete sequence of amino acids representing the unknown peptide or protein,
comparing the complete sequence of amino acids to a reference proteome, and
identifying a known peptide or protein based on the comparison.
12. The system of claim 11 , wherein the preprocessing comprises using gel electrophoresis and/or liquid chromatography to isolate the unknown peptide or protein from the mixture of proteins and molecules, and wherein the fixed-dimensional vector representation is an m-dimensional or one-dimensional vector of m elements, where m corresponds to a number of different m/z values and abundances for the n-most abundant ions, and wherein the decoding comprises mapping the m-dimensional or one-dimensional vector of m elements back to the sequence of amino acids while taking into consideration a set of metadata features and a context vector.
13. The system of claim 11 , wherein the identifying further comprises:
matching one or more portions of the complete sequence of amino acids to one or more corresponding portions of a known peptide sequence;
determining a probability that the match between the one or more portions of the complete sequence of amino acids and the one or more corresponding portions of the known peptide sequence is indicative of a presence of the known peptide sequence in the sample; and
when the probability is greater than a predetermined threshold, obtaining an identifier of the corresponding peptide or protein sequence from the reference proteome,
wherein the unknown peptide or protein is identified based on the identifier.
14. The system of claim 11 , wherein each long short term memory cell comprises three gates and a cell unit, wherein each of the three gates is configured to use a sigmoid activation, and wherein a cell state is determined by transforming data associated with at least one gate using a hyperbolic tangent function.
15. The system of claim 11 , wherein each of the first and the second bidirectional recurrent neural network comprises at least two sets of hidden layers, wherein each set of the hidden layers is configured to processing data in a different direction.