IBSB tutorial
Download as PPTX, PDF0 likes206 views
This document discusses computational methods for identifying metabolites from tandem mass spectrometry data. It begins with background on metabolites and challenges in identification. Common approaches are described, including mass spectra libraries, in silico fragmentation using rules or machine learning, and machine learning methods. Recent machine learning works are summarized, such as using kernels to model peak interactions, unsupervised methods to group metabolites by shared substructures, and automatically recommending substructures from mass spectra. The document concludes that metabolite identification is important for metabolomics and machine learning is key to recent advances.
![Machine learning Approach
Supervised ML for substructure prediction
SIMPLE (Bioinformatics, 2018)
• Idea: introducing interaction term to model (two-way interaction model)
• Prediction model:
𝑓 𝑥 = 𝑏 + 𝑤 𝑇 𝑥 + 𝑥 𝑇 𝑊𝑥 , 𝑦 𝑥 = 𝑠𝑔𝑛(𝑓(𝑥))
• Objective function :
min
𝑏,𝑤,𝑊
𝑖=1
𝑛
[1 − 𝑦𝑖 𝑓(𝑥𝑖)]+ + 𝛼 𝑤 1 + 𝛽 𝑊 ∗
• Convexity guarantees to find globally optimal solution.
Hinge loss Sparsity Low-rank
Peaks Interactions
20/07/2018 D. H. Nguyen, Kyoto University 18](https://image.slidesharecdn.com/ibsbtutorial-180720001942/85/IBSB-tutorial-18-320.jpg)
![SIMPLE (Bioinformatics, 2018)
Idea: use background knowledge (interactions from trees) to regularize W.
Laplacian regularization
𝑥 𝑇 𝑊𝑥 = 𝑖,𝑗 𝑤𝑖𝑗 𝑥𝑖 𝑥𝑗 = 𝑖,𝑗(𝑣𝑖
𝑇
𝑣𝑗)𝑥𝑖 𝑥𝑗
𝑊 can be decomposed as 𝑉 𝑇
𝑉 (low rank decomposition)
𝑅 𝑉 = 𝑖,𝑗 𝐴𝑖𝑗 𝑣𝑖 − 𝑣𝑗
2
= trace 𝑊𝐿 ,
where 𝐿 is Laplacian matrix.
New objective function :
min
𝑏,𝑤,𝑊
𝑖=1
𝑛
[1 − 𝑦𝑖 𝑓(𝑥𝑖)]+ + 𝛼 𝑤 1 + 𝛽 𝑊 ∗ + 𝛾 trace(𝑊𝐿)
Still convex
Machine learning Approach
Supervised ML for substructure prediction
20/07/2018 D. H. Nguyen, Kyoto University 19
+](https://image.slidesharecdn.com/ibsbtutorial-180720001942/85/IBSB-tutorial-19-320.jpg)
IBSB tutorial
- 1. Computational methods for metabolite identification from tandem mass spectrometry Dai Hai Nguyen Kyoto University, Japan 20/07/2018 D. H. Nguyen, Kyoto University 1
- 2. Background of metabolites identification Metabolites Intermediate or end products of metabolism Small molecules with important functions: energy transport, building blocks of cells, etc. Many applications, e.g. drug discovery Identifying or profiling them is challenging 20/07/2018 D. H. Nguyen, Kyoto University 2
- 3. Background of metabolites identification Tandem Mass spectrometry fragments compound into many fragments each fragment corresponds to a peak There exist peak interactions (co-occurrence of peaks) 20/07/2018 D. H. Nguyen, Kyoto University 3 Peak interaction
- 4. Background of metabolites identification Task: given a query spectrum, find similar molecules in database. Approaches: 20/07/2018 D. H. Nguyen, Kyoto University MS library In silico fragmentation Machine learning 4
- 5. I. Mass spectra library Simply compare query spectrum with spectra in library Best matching candidates are returned Drawback: size of library is limited E.g., Human metabolome database ~ 2000 compounds 20/07/2018 D. H. Nguyen, Kyoto University MS library 5
- 6. II. In silico fragmentation To mitigate insufficiency of spectra library by taking advantage of structural database. Can be divided into groups: 1) rule-based 2) combinatorial based 3) machine learning based 20/07/2018 D. H. Nguyen, Kyoto University In silico fragmentation 6
- 7. II. In silico fragmentation (1) 1) Rule based fragmentation, e.g., Mass Frontier Use set of fragmentation rules to predict spectra from compound structures. Rules are extracted from the literature. Not preferred in practice due to: fragmentation process can be variant due to small changes in molecular structure # rules insufficient to identify fragments with high accuracy intensities of peaks are ignored 20/07/2018 D. H. Nguyen, Kyoto University 7
- 8. II. In silico fragmentation (2) 2) Combinatorial based fragmentation, e.g. FiD From molecular structure, generate graph of all connected substructures. Find most likely fragmentation trees that best matches spectrum. Drawbacks: computationally expensive -> applied for small molecules Intensities of fragments are ignored 20/07/2018 D. H. Nguyen, Kyoto University 8
- 9. II. In silico fragmentation (3) 3) Machine learning based fragmentation Use ML to learn fragmentation process from data. Peak intensities are considered and learned Very few work 20/07/2018 D. H. Nguyen, Kyoto University 9
- 10. II. In silico fragmentation (3) Competitive Fragmentation Modeling (CFM) models fragmentation as a Markov process of state transition between fragments 1. Transition model 2. Observation model 20/07/2018 D. H. Nguyen, Kyoto University 10
- 11. Background of metabolites identification Task: given a query spectrum, find similar molecules in database. Approaches: 20/07/2018 D. H. Nguyen, Kyoto University MS library In silico fragmentation Machine learning 11
- 12. III. Machine learning Approach a) Supervised ML for substructure prediction b) Unsupervised ML for substructure annotation 20/07/2018 D. H. Nguyen, Kyoto University 12
- 13. IV. Machine learning Approach supervised ML for substructure prediction Step 1: fingerprint prediction Step 2: Candidate retrieval 20/07/2018 D. H. Nguyen, Kyoto University 13
- 14. Machine learning Approach Supervised ML for substructure prediction FingerID (Bioinformatics, 2012) Kernel method • Define probability product kernel (PPK) for spectra. • Then, use SVM for classification. Drawback Peak interactions are ignored. Limited accuracy 𝑝 𝑋 = 1 𝑛 𝑋 𝑘=1 𝑛 𝑋 𝑝 𝑋(𝑘) 𝑝 𝑌 = 1 𝑛 𝑌 𝑘=1 𝑛 𝑌 𝑝 𝑌(𝑘) 𝐾 𝑋, 𝑌 = 1 𝑛 𝑋 𝑛 𝑌 𝑖,𝑗 𝑝 𝑋(𝑖)𝑝 𝑌(𝑗) 20/07/2018 D. H. Nguyen, Kyoto University 14
- 15. Machine learning Approach Supervised ML for substructure prediction CSI:FingerID (Bioinformatics, 2014) Improved version of FingerID Define kernel for spectra by PPK Kernels for fragmentation trees are defined and combined with PPK via MKL. Then, use SVM for classification. 20/07/2018 D. H. Nguyen, Kyoto University 15
- 16. Machine learning Approach Supervised ML for substructure prediction CSI:FingerID (Bioinformatics, 2014) Fragmentation trees Models of fragmentation of a molecule in MS/MS Nodes ~ peaks ~ molecular formula of fragments. Edges ~ losses ~ uncaptured uncharged fragments. Trees can be predicted from spectra provide structural information of spectra. 20/07/2018 D. H. Nguyen, Kyoto University 16
- 17. Machine learning Approach Supervised ML for substructure prediction CSI:FingerID (Bioinformatics, 2014) Pros & Cos Improved accuracy due to additional structural information provided by trees Computationally expensive due to conversion of trees from spectra Lack of interpretation 20/07/2018 D. H. Nguyen, Kyoto University 17
- 18. Machine learning Approach Supervised ML for substructure prediction SIMPLE (Bioinformatics, 2018) • Idea: introducing interaction term to model (two-way interaction model) • Prediction model: 𝑓 𝑥 = 𝑏 + 𝑤 𝑇 𝑥 + 𝑥 𝑇 𝑊𝑥 , 𝑦 𝑥 = 𝑠𝑔𝑛(𝑓(𝑥)) • Objective function : min 𝑏,𝑤,𝑊 𝑖=1 𝑛 [1 − 𝑦𝑖 𝑓(𝑥𝑖)]+ + 𝛼 𝑤 1 + 𝛽 𝑊 ∗ • Convexity guarantees to find globally optimal solution. Hinge loss Sparsity Low-rank Peaks Interactions 20/07/2018 D. H. Nguyen, Kyoto University 18
- 19. SIMPLE (Bioinformatics, 2018) Idea: use background knowledge (interactions from trees) to regularize W. Laplacian regularization 𝑥 𝑇 𝑊𝑥 = 𝑖,𝑗 𝑤𝑖𝑗 𝑥𝑖 𝑥𝑗 = 𝑖,𝑗(𝑣𝑖 𝑇 𝑣𝑗)𝑥𝑖 𝑥𝑗 𝑊 can be decomposed as 𝑉 𝑇 𝑉 (low rank decomposition) 𝑅 𝑉 = 𝑖,𝑗 𝐴𝑖𝑗 𝑣𝑖 − 𝑣𝑗 2 = trace 𝑊𝐿 , where 𝐿 is Laplacian matrix. New objective function : min 𝑏,𝑤,𝑊 𝑖=1 𝑛 [1 − 𝑦𝑖 𝑓(𝑥𝑖)]+ + 𝛼 𝑤 1 + 𝛽 𝑊 ∗ + 𝛾 trace(𝑊𝐿) Still convex Machine learning Approach Supervised ML for substructure prediction 20/07/2018 D. H. Nguyen, Kyoto University 19 +
- 20. Machine learning Approach Supervised ML for substructure prediction Input Output Kernel Regression (IOKR) (Bioinformatics, 2017) Idea: using IOKR to learn the mapping between spectra and molecular structure. Two steps: 1. Estimation of the output feature map by solving 2. Computation of pre-image problem 20/07/2018 D. H. Nguyen, Kyoto University 20
- 21. Machine learning Approach Unsupervised ML for substructure annotation Metabolites/molecules may have common substructures, yielding similar fragments/peaks in spectra. Such substructures are pertaining to biochemical processes Allows to group metabolites based on shared substructures Improve the accuracy of metabolite identification 20/07/2018 D. H. Nguyen, Kyoto University 21
- 22. IV. Machine learning Approach Unsupervised ML for substructure annotation MS2LDA (Bioinformatics 2017) Automatically extract relevant substructures in molecules in metabolites based on co-occurrence of fragments and losses. Motivated by topic modeling for text application. e.g. Latent Dirichlet Allocation (LDA) LDA for MS data (MS2LDA) Peaks ~ words set of peaks (substructures) ~ topics LDA decompose a text into topics, while MS2LDA decomposes a molecule into substructures. Drawbacks: extracted substructures need to be annotated based on expert knowledge (complex process and time- consuming) 20/07/2018 D. H. Nguyen, Kyoto University 22
- 23. Machine learning Approach Unsupervised ML for substructure annotation Automated recommendation of subtructures from MS/MS (Aida Mrzic et al, bioRxiv) Automatically extract relevant substructures in molecules based on co-occurrence of fragments and losses Applied Frequent Itemset Mining to extract association rules. Given query spectrum, get recommendation of substructures present in it by applying extracted rules. 20/07/2018 D. H. Nguyen, Kyoto University 23
- 24. Conclusion • Metabolite Identification is an essential part in metabolomics to enlarge knowledge of biological systems. • Many techniques/software with different approaches have been proposed to deal with this task and can be categorized into groups • ML methods are the key to recent progress in metabolite identification 20/07/2018 D. H. Nguyen, Kyoto University 24
Editor's Notes
- #2: 大海