A method for predicting a molecular fingerprint of a compound
By constructing a vector representation of mass spectrometry data, incorporating neutral loss information, and utilizing natural language processing techniques, the problem of insufficient coverage in compound identification during mass spectrometry data analysis was solved, achieving more efficient compound qualitative analysis and molecular fingerprint prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- DALIAN UNIV OF TECH
- Filing Date
- 2023-07-13
- Publication Date
- 2026-05-29
AI Technical Summary
Existing mass spectrometry data analysis methods have insufficient coverage in compound identification, especially when compounds are not in the database. Furthermore, existing methods fail to fully utilize the implicit information in mass spectrometry data.
By constructing a vector representation of mass spectrometry data, incorporating neutral loss information, and utilizing natural language processing techniques to learn the relationships between peaks and between peaks and neutral loss, machine learning or deep learning methods are employed to predict the molecular fingerprint of compounds.
It improves the accuracy of compound qualitative analysis and enhances the accuracy and efficiency of molecular fingerprint prediction. In particular, when using Word2Vec to train mass spectrometry data, the generated spectral vectors contain richer structural information, thus improving the compound identification effect.
Smart Images
Figure CN117059174B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mass spectrometry data analysis technology, specifically relating to a method for predicting molecular fingerprints using mass spectrometry data. Background Technology
[0002] Metabolite annotation is a crucial and challenging step in mass spectrometry-based metabolomics analysis. Mass spectrometry is a widely used technique for identifying compounds in biological systems. Matching the mass spectrum to be detected with a database of mass spectra, calculating the similarity between the unknown mass spectrum and standard mass spectra in the database, and identifying the standard structure with the highest similarity as the unknown structure has become a standard method for metabolite annotation. Despite its widespread use, this method suffers from a coverage problem: if the compound being searched for is not in the database, it cannot be identified.
[0003] To address this issue, researchers have proposed numerous methods, such as calculating the theoretical mass spectra of compounds based on their chemical structures, or predicting mass spectra of compounds using molecular fingerprints through neural networks. The former is primarily rule-based, while the latter mainly utilizes quantum chemistry and machine learning methods. However, both methods require substantial expertise, and the application of quantum chemistry methods demands high computational power. Machine learning methods primarily predict the probability of bond breaking in molecular structures under experimental conditions; the most representative method is CFM-ID, used to predict ESI-MS / MS mass spectra. For predicting mass spectra using molecular fingerprints, the most representative method is NEIMS, which can directly predict EI-MS mass spectra from molecular fingerprints. Compared to CFM-EI, NEIMS not only has better prediction accuracy but also runs faster. Furthermore, the MS / MS mass spectra of unknown metabolites can be compared with the mass spectra of reference metabolites, and their structural similarity can be predicted by calculating the similarity score between the two for identification, such as the DeepMASS method, which uses MS / MS mass spectra and deep neural networks to score the structural similarity between unknown and known metabolites.
[0004] Besides the methods mentioned above, other compound identification strategies exist. For example, to rank candidate structures based on obtained mass spectra, molecular fingerprints can be predicted from the mass spectra. These predicted fingerprints are then compared with the fingerprints of compounds in a structure database to rank the candidate compounds. This method is generally used in combination with other methods to achieve better performance. MetExpert was developed to predict molecular fingerprints from EI-MS mass spectra and combined with retention index prediction for compound similarity assessment. However, MetExpert only uses 188-bit MACCS fingerprints, resulting in low recognition rates for similar structures. MetExpert uses a PLS-DA (Partial Least Squares Discriminant Analysis) model for molecular fingerprint prediction. Additionally, in 2020, a deep neural network model, DeepEI, was proposed to directly retrieve the structure of unknown compounds from a molecular structure database using EI-MS mass spectra. This method uses a deep neural network to predict multiple molecular fingerprints from EI-MS mass spectra (FP model). Then, the predicted molecular fingerprints are used to search a molecular structure database to retrieve the structure of the unknown compound.
[0005] However, the above methods also have certain drawbacks. Mass spectrometry data contains rich structural information about compounds. Most studies only discretize mass spectrometry data to convert it into a corresponding vector representation, utilizing only the information of the peaks. Other implicit information in mass spectrometry data, such as neutral loss, is not fully utilized. This invention, from the perspective of fully mining the implicit information of mass spectrometry data, adds neutral loss information after the corresponding ion peaks when constructing the vector representation of mass spectrometry data. Using natural language processing technology, it learns the relationships between peaks and between peaks and neutral loss in the mass spectrum, constructing a multi-dimensional spectral vector to predict molecular fingerprints. Due to the addition of neutral loss of the corresponding ion peaks in the spectral vector, richer structural information is contained, which can effectively improve the accuracy of molecular fingerprint prediction. Summary of the Invention
[0006] The purpose of this invention is to construct a vector representation of mass spectrometry data, reasonably incorporate neutral loss into the spectral vector, and utilize Word2vec technology in natural language to more accurately learn the potential relationships between peaks and between peaks and neutral loss, so that the spectral vector carries richer structural information. Based on the spectral vector, classification algorithms using machine learning or deep learning are used to predict the molecular fingerprint of compounds, thereby better assisting in the qualitative analysis of compounds.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] The method consists of two parts. First, the mass spectrometry data is preprocessed to construct an effective vector representation of the mass spectra, ensuring it contains rich structural information about the compounds. A mass spectrum can be represented by a set of mass peaks. As shown in the figure (from the MassBank database), the following is the structural formula of the molecule:
[0009]
[0010] Current machine learning or deep learning algorithms for predicting molecular fingerprints merely vectorize mass spectrometry data, converting it into a multi-dimensional vector using the mass number and peak intensity of ion peaks. This ignores other structural information implicit in the mass spectrum, such as neutrality loss. Constructing a better representation of mass spectrometry data that includes more structural information can improve the accuracy of molecular fingerprint prediction, thereby enhancing the accuracy of compound identification. A molecular fingerprint represents the substructure information of a compound, specifically represented as a column vector of 0s and 1s, where each bit represents the presence or absence of a specific chemical structure. The rich information in mass spectrometry makes it possible to predict molecular fingerprints directly from mass spectra. A detailed description of the experimental research is as follows:
[0011] To obtain a more effective representation of mass spectra, it is necessary to make full use of the information contained in the mass spectra peaks and the implicit neutral loss information. Here, we draw on the Spec2Vec method proposed by Florian Huber et al. in 2021 to preprocess the compound mass spectrometry data. This method treats the compound mass spectrum as an article, converting the mass spectra peaks and neutral loss into textual form, learning the vector representations of the mass spectra peaks and neutral loss, and thus obtaining the vector representation of the mass spectrometry data. This technique requires a series of processing steps on the mass spectrometry data:
[0012] 1) Normalize the peak intensity in each mass spectrum.
[0013] 2) Remove peaks with m / z greater than 1000 DA from each mass spectrum.
[0014] 3) Add a neutral loss in the range of 5~1000DA.
[0015] Next, the processed mass spectra are converted into documents. Each peak is converted into a word containing its corresponding position; for example, the peak at m / z 100.45 is converted into the word "peak@100.45". When converting the calculated neutral loss into words, since the mass number of the neutral loss fragments is obtained by subtracting the mass of the corresponding fragment ion from the mass of the parent ion, the neutral losses obtained from different secondary mass spectra may have the same mass number but represent different structures. To eliminate the influence of the above situation, in the process of converting neutral losses into words, neutral losses for some specific structures are converted into the form of loss@xxx.xx, and the rest are converted into the form of loss@(parent ion mass - fragment mass). Mass spectrum peaks and neutral losses are essentially fragments of compounds after being broken down. In order to better reflect the relationship between peaks and neutral losses, this invention pairs the formed words such as peak@xxx.xx and loss@xxx.xx according to the m / z after @, thereby converting the mass spectrum into a document composed of words. Word2Vec is used to train the mass spectrum document to obtain the vector representation of each peak. Word2Vec learns the relationships between words from all documents, representing words through abstract word vectors (word embeddings) to place words with similar meanings close together. A mass spectrum consists of multiple peaks, so the spectral vector is represented by the sum of the word vectors it contains. To illustrate the correlation between peaks and peak intensities, a weighted summation is used to calculate the vector corresponding to each spectrum, as shown below:
[0016]
[0017] in This represents the obtained spectral vector. This represents the vector corresponding to each peak and neutral loss obtained after Word2Vec training. This represents the corresponding intensity.
[0018] Compared to methods that discretize mass spectra to form spectral vectors, this method generates spectral vectors containing richer information, which helps improve the accuracy of molecular fingerprint prediction.
[0019] The second part is molecular fingerprint prediction, which predicts the molecular fingerprint of a compound based on the mass spectrometry vector obtained in the first part. A molecular fingerprint is a multi-dimensional vector, where each bit consists of either 0 or 1, representing the presence or absence of a specific chemical structure. Predicting each bit of the molecular fingerprint is equivalent to performing a binary classification task, thus allowing the use of machine learning algorithms and deep learning methods. A prediction model is trained for each bit of the molecular fingerprint to represent the substructure information of the unknown compound. The overall workflow is as follows: Figure 2 As shown. Attached Figure Description
[0020] Figure 1 This is the mass spectrum of the molecule. The horizontal axis represents the mass-to-charge ratio (m / z), and the vertical axis represents the peak intensity.
[0021] Figure 2 This is a flowchart of the process of the present invention. Detailed Implementation
[0022] (1) Data source: The experiment was conducted using the LC-MS dataset provided on GNPS. Each mass spectrometry data consists of important information such as molecular formula, fragment peak mass and peak intensity.
[0023] (2) Data processing
[0024] For each mass spectrometry data point, its corresponding SMILES is obtained using the chemical formula provided in the database. SMILES is a specification that explicitly describes molecular structures using ASCII strings. To ensure that each SMILES corresponds to a unique chemical structure, this invention uses Canonical Smiles. After excluding duplicate mass spectra, compounds with molecular weights exceeding 2000 Da, and mass spectra containing uncommon elements (excluding C, H, O, N, P, S, F, Cl, Br, and Si), 12797 mass spectra were retained. Simultaneously, their corresponding molecular fingerprints were obtained using the SMILES. The training and test sets were split in a 9:1 ratio.
[0025] (3) Convert mass spectrometry data into vectors
[0026] Word2vec is used to represent each mass spectrometry data processed in (1) as a 300-dimensional vector. When training the Word2Vec model, the CBOW model is selected and the window size is set to 3. This parameter setting can better learn the relationship between peaks and neutral loss in the mass spectrometry data, so that the obtained spectral vector can be better used for subsequent molecular fingerprint prediction.
[0027] (4) Predicting molecular fingerprints
[0028] The abundant fragments in ESI-MS mass spectrometry make it possible to predict molecular fingerprints directly from mass spectra. We train a classifier model for each fingerprint, using MLP, LR, PLS-DA, and XGBoost methods to predict each molecular fingerprint. First, we examine the impact of Word2vec training with spectral vector neutral loss on molecular fingerprint prediction; detailed results are shown in Table 1.
[0029] Table 1 Comparison of the impact of neutral loss on molecular fingerprint prediction performance
[0030]
[0031] Table 1 shows the role of neutral loss information in molecular fingerprint prediction. This indicates that neutral loss information was added when training mass spectrometry data using Word2Vec to obtain spectral vectors. The results in the table show that adding the implicit neutral loss information in mass spectrometry during the generation of spectral vectors improved the performance of various methods.
[0032] When using neutral loss information, changes were made compared to the Spec2Vec method. The impact of the two methods on molecular fingerprint prediction results is now compared, and detailed results are shown in Table 2.
[0033] Table 2 compares the impact of neutral loss addition location on molecular fingerprint prediction performance.
[0034]
[0035] As shown in Table 2, adjusting the position of words transformed by neutral loss makes the generated spectral vector more suitable for molecular fingerprint prediction. This indicates that the position of the neutral loss was adjusted when training mass spectrometry data using Word2Vec to obtain spectral vectors. The results in the table show that the performance of each method was improved after the adjustment.
[0036] Table 3 Compares the impact of neutral loss generation methods on molecular fingerprint prediction performance.
[0037]
[0038] As shown in Table 3, by adjusting the generation method of neutral loss, the generated spectral vectors are made more suitable for molecular fingerprint prediction. This indicates that the method for generating the neutral loss was adjusted when using Word2Vec to train mass spectrometry data to obtain spectral vectors. As can be seen from the results in the table, the accuracy and precision of each method have been improved to a certain extent.
[0039] Table 4 compares the impact of different spectral vector generation methods on molecular fingerprint prediction performance.
[0040]
[0041] Table 4 shows a comparison between molecular fingerprint prediction using the spectral vectors obtained after training with Word2Vec and molecular fingerprint prediction using traditional input methods. Compared to traditional input methods, LR, PLS-DA, and XGBoost methods are slightly inferior in accuracy. However, in MLP methods, the combination of pre-training shows advantages across all metrics.
[0042] To better compare the performance differences of the various methods and illustrate the advantages of combining preprocessing with neural networks, the results of each method achieving advantages in various performance aspects are listed in Table 1.
[0043] Table 5. Performance comparison of preprocessing combined with MLP method and other methods.
[0044]
[0045] Table 5 illustrates the performance advantages of using the spectral vectors trained by this technique as input and employing a deep neural network as the molecular fingerprint prediction model. Compared to the PLS-DA method, it is slightly inferior in prediction accuracy, but superior in all other metrics.
[0046] Since the preserved molecular fingerprints are predicted separately, the overall distribution of the results should be more even and reasonable, rather than polarized. Table 6 shows the distribution of various methods on ACC and PRE.
[0047] Table 6. Distribution of results among multiple methods
[0048]
Claims
1. A method for predicting molecular fingerprints of compounds, characterized in that, It is divided into two parts: the first part is to obtain the mass spectrometry vector; the second part is to predict the molecular fingerprint of the compound. Part 1: Obtaining the Mass Spectrum Vector First, the mass spectrometry data is preprocessed to construct an effective vector representation of the mass spectrometry, which contains rich structural information of the compounds. A mass spectrum can be represented by a set of mass peaks; The Spec2Vec method is used to preprocess compound mass spectrometry data. This method treats the compound mass spectrum as an article, converting mass peaks and neutral loss into textual form, learning vector representations of mass peaks and neutral loss, and thus obtaining a vector representation of the mass spectrometry data. A series of processing steps are then performed on the mass spectrometry data: 1) Normalize the peak intensity in each mass spectrum; 2) Remove peaks with m / z greater than 1000 DA from each mass spectrum; 3) Add a neutral loss in the range of 5~1000 DA; The processed mass spectrometry data is converted into documents. Each peak is transformed into a word containing its corresponding position. During the conversion of neutral loss into words, for neutral loss of a specific structure, it is converted into the form `loss@xxx.xx`, while other neutral losses are converted into `loss@parent ion mass - fragment mass`. Mass spectrometry peaks and neutral losses are fragments of the compound after fragmentation. The resulting words `peak@xxx.xx` and `loss@xxx.xx` are paired according to the `m / z` after `@`, thus converting the mass spectrometry data into documents composed of words. Word2Vec is used to train the mass spectrometry documents, obtaining a vector representation of each peak. Word2Vec learns the relationships between words from all documents, representing words through abstract word vectors to place words with similar meanings close together. A mass spectrometry spectrum consists of multiple peaks, so the spectral vector is represented by the sum of the word vectors it contains. To illustrate the correlation between peaks and peak intensities, a weighted summation method is used to calculate the vector corresponding to each spectrum, as shown below: in This represents the obtained spectral vector. This represents the vector corresponding to each peak and neutral loss obtained after Word2Vec training. This represents the corresponding intensity; Part Two: Prediction of Molecular Fingerprints Based on the mass spectrometry vector obtained in the first part, the molecular fingerprint of the compound is predicted. The molecular fingerprint is a multi-dimensional vector, each of which is composed of 0 or 1, representing the presence or absence of a certain chemical structure. The prediction of each molecular fingerprint is equivalent to performing a binary classification task. Machine learning algorithms and deep learning methods are used for prediction. A prediction model is trained for each of the molecular fingerprints to represent the substructure information of the unknown compound.
2. The method for predicting molecular fingerprints of compounds according to claim 1, characterized in that, In the first part, the Word2Vec technique from natural language processing is used to convert mass spectrometry data into documents. At the same time, the relationship between peaks and neutral loss in the mass spectrometry data is considered to learn the structural relationships of compounds hidden in the mass spectrometry data, so as to obtain a more reasonable representation of the mass spectrometry data and thus better apply it to subsequent molecular fingerprint prediction.
3. A method for predicting molecular fingerprints of compounds according to claim 1 or 2, characterized in that, In the second part, the methods used to predict molecular fingerprints are MLP, LR, PLS-DA, and XGBoost.