Recommended
More Related Content
Similar to Segway and the Graphical Models Toolkit: a framework for probabilistic genomic inference
Similar to Segway and the Graphical Models Toolkit: a framework for probabilistic genomic inference (20)
Recently uploaded
Recently uploaded (20)
Segway and the Graphical Models Toolkit: a framework for probabilistic genomic inference
- 1. Michael M. Hoffman Princess Margaret Cancer Centre Vector Institute Department of Medical Biophysics Department of Computer Science University of Toronto https://hoffmanlab.org/ Segway and the Graphical Models Toolkit: A framework for probabilistic genomic inference @michaelhoffman #probgen18
- 2. Geographical maps have… Figures from 1) National Geographic Society (2011) GIS. street data + buildings data + vegetation data = integrated map Rachel Chan
- 3. Functional genomics ENCODE Project Consortium 2011. PLoS Biol 9:e1001046.
- 4. Functional genomics ENCODE Project Consortium 2011. PLoS Biol 9:e1001046.
- 5. Functional genomics ENCODE Project Consortium 2011. PLoS Biol 9:e1001046.
- 6. Functional genomics ENCODE Project Consortium 2011. PLoS Biol 9:e1001046.
- 7. Semi-automated genome annotation genomic signal pattern discovery visualization annotation interpretation
- 8. Segway A way to segment the genome https://segway.hoffmanlab.org/ Hoffman MM et al. 2012. Nat Methods 9:473.
- 11. 0 1 0 21 1 Finite number of labels
- 12. 0 1 0 1 1 Maximize similarity in labels 2
- 13. 0 1 0 1 1 Maximize similarity in labels 2
- 14. 0 1 0 1 1 Maximize similarity in labels 2
- 15. TSS transcription start site GS gene start GM gene middle GE gene end E enhancer I distal CTCF R repression D dead
- 16. Transcription start site (TSS) Hoffman MM et al. 2013. Nucleic Acids Res 41:827.
- 17. Graphical Models Toolkit (GMTK) • General purpose toolkit for probabilistic inference in temporal signals (speech, language, activity recognition, genomics) • Written in highly optimized scalable C++ • Supports many machine learning algorithms, inference procedures, and probability models. • Can express arbitrary structured dynamic graphical models • http://melodi.ee.washington.edu/gmtk • conda install -c bioconda gmtk Jeff Bilmes
- 19. Q hidden random variable discrete P(q|θ) = P(Q = q) Equation Graphical realization parameters
- 20. Q hidden random variable discrete P(q|θ) = P(Q = q) variable: state { type: discrete hidden cardinality 2; conditionalparents: nil using DenseCPT("start_state"); } Equation Graphical GMTKL
- 21. Q hidden random variable discrete P(q|θ) = P(Q = q) variable: state { type: discrete hidden cardinality 2; conditionalparents: nil using DenseCPT("start_state"); } DENSE_CPT_IN_FILE inline % conditional probability tables 1 % total number of CPTs = 1 0 start_state % CPT #0, "start_state" 0 % 0 parents 2 % output cardinality 2 0.25 % P(Q = 0) = 0.25 0.75 % P(Q = 1) = 0.75 Equation Graphical GMTKL Parameters
- 22. Q X hidden random variable conditional relationship observed random variable discrete continuous P(q,x|θ) = P(Q = q)P(X = x|Q = q) variable: state { type: discrete hidden cardinality 2; conditionalparents: nil using DenseCPT("start_state"); } variable: obs { type: continuous observed 0:0; conditionalparents: state(0) using mapping("state_obs"); }
- 23. Q0 X0 hidden random variable conditional relationship observed random variable discrete continuous P(q0,x0|θ) = P(Q₀ = q₀)P(X₀ = x₀|Q₀ = q₀) frame: 0 { variable: state { type: discrete hidden cardinality 2; conditionalparents: nil using DenseCPT("start_state"); } variable: obs { type: continuous observed 0:0; conditionalparents: state(0) using mapping("state_obs"); } } X0
- 24. Q0 X0 hidden random variable conditional relationship observed random variable discrete continuous Q1 P(q0:1,x0|θ) = P(Q₀ = q₀)P(X₀ = x₀|Q₀ = q₀)P(Q₁ = q₁|Q₀ = q₀) frame: 0 { variable: state { type: discrete hidden cardinality 2; conditionalparents: nil using DenseCPT("start_state"); } variable: obs { type: continuous observed 0:0; conditionalparents: state(0) using mapping("state_obs"); } } frame: 1 { variable: state { type: discrete hidden cardinality 2; conditionalparents: state(-1) using DenseCPT("state_state"); } }
- 25. Q0 X0 hidden random variable conditional relationship observed random variable discrete continuous Q1 X1 P(q0:1,x0:1|θ) = P(Q₀ = q₀)P(X₀ = x₀|Q₀ = q₀)P(Q₁ = q₁|Q₀ = q₀)P(X₁ = x₁|Q₁ = q₁) frame: 0 { variable: state { type: discrete hidden cardinality 2; conditionalparents: nil using DenseCPT("start_state"); } variable: obs { type: continuous observed 0:0; conditionalparents: state(0) using mapping("state_obs"); } } frame: 1 { variable: state { type: discrete hidden cardinality 2; conditionalparents: state(-1) using DenseCPT("state_state"); } variable: obs { type: continuous observed 0:0; conditionalparents: state(0) using mapping("state_obs"); } }
- 26. Xt+1 hidden random variable conditional relationship observed random variable discrete continuous frame: 0 { variable: state { type: discrete hidden cardinality 2; conditionalparents: nil using DenseCPT("start_state"); } variable: obs { type: continuous observed 0:0; conditionalparents: state(0) using mapping("state_obs"); } } frame: 1 { variable: state { type: discrete hidden cardinality 2; conditionalparents: state(-1) using DenseCPT("state_state"); } variable: obs { type: continuous observed 0:0; conditionalparents: state(0) using mapping("state_obs"); } } chunk 1:1 Qt XtXt Qt+1 Xt+1 P(q0:T,x0:T|θ) = P(Q₀ = q₀)P(X₀ = x₀|Q₀ = q₀) Πt=1 P(Qt = qt|Qt-1 = qt-1)P(Xt = xt|Qt = qt) T
- 27. hidden random variable conditional relationship observed random variable discrete continuous Qt Xt Xt+1 Qt+1 Xt+2 Qt+2 Xt+3 Qt+3 Xt+4 Qt+4 Xt+5 Qt+5 Dynamic Bayesian network (hidden Markov model) P(q0:T,x0:T|θ) = P(Q₀ = q₀)P(X₀ = x₀|Q₀ = q₀) Πt=1 P(Qt = qt|Qt-1 = qt-1)P(Xt = xt|Qt = qt) T
- 28. Dynamic Bayesian Network for segmentation segment label CTCF H3K36me3 DNaseI hidden random variable conditional relationship observed random variable discrete continuous
- 29. Segway work flow .fasta.gz genomedata-load segway train .genomedata .bed.gz .bedGraph.gz .wig.gz segway annotate GMTKL structure starting params discovered params segway.bed.gz segtools .png, .pdf, .tab
- 30. What does Segway do, really? 1. Creates GMTKL structure and parameter files for semi- automated genome annotation 2. Converts genomic data to a GMTK binary observation format 3. Runs GMTK to perform EM training, Viterbi decoding, posterior inference 4. Manages job execution via cluster (SGE, LSF, PBS, Slurm), or local multiprocessing 5. Converts GMTK output to genomics formats (BED, wiggle)
- 31. Segway modular API (pull request #91) segway train segway train-init segway train-run segway train-run-round segway train-finish segway annotate segway annotate-init segway annotate-run segway annotate-finish segway posterior segway posterior-init segway posterior-run segway posterior-finish
- 32. Handling missing data hidden random variable conditional observed random variable segment DNaseI discrete continuous switching 1 0
- 33. segment label CTCF H3K36me3 DNaseI hidden random variable conditional observed random variable discrete continuous switching present(CTCF) present(H3K36me3) present(DNaseI) Handling missing data
- 35. segment label CTCF H3K36me3 DNaseI present(CTCF) present(H3K36me3) present(DNaseI) segment countdown segment transition rulerframe indexLength distribution • Minimum segment length • Maximum segment length • Trained geometric length distribution • Dirichlet prior on segment length • Weight of prior versus observed data
- 36. Hierarchical models • Two-level model topology • Mixture of large- scale and punctate labels “closed” “open” “intermediate”
- 38. segment label long RNA-seq cytosol short RNA-seq whole cell CAGE nucleus + strand hidden random variable conditional relationship observed random variable discrete continuous
- 39. segment label long RNA-seq cytosol short RNA-seq whole cell CAGE nucleus + strand – strand hidden random variable conditional relationship observed random variable discrete continuous
- 42. Adding evolution segment Ptro nucleotide Ptro dinucleotide Hsap nucleotide H3K27me3 Hsap dinucleotide
- 43. Adding variation segment YRI nucleotide YRI dinucleotide ancestral nucleotide YRI DAF ancestral dinucleotide
- 47. Acknowledgments The Hoffman Lab Jeff Bilmes William Noble Max Libbrecht Funding: Princess Margaret Cancer Foundation Canadian Institutes of Health Research Canadian Cancer Society Natural Sciences and Engineering Research Council Ontario Institute for Cancer Research Ontario Ministry of Research, Innovation and Science Medicine by Design McLaughlin Centre Samantha Wilson Coby Viner Mickaël Mendez Danielle Denisko Chang Cao Lee Zamparo Eric Roberts Mehran Karimzadeh Francis Nguyen Rachel Chan Matthew McNeil Natalia Mukhina
- 48. Postdoctoral, MSc, PhD positions available in my research lab at the Princess Margaret Cancer Centre Dept of Medical Biophysics Dept of Computer Science University of Toronto Please approach me for details. Michael Hoffman https://hoffmanlab.org/ michael.hoffman@utoronto.ca @michaelhoffman
Editor's Notes
- - need laser pointer - turn Workrave off turn phone off turn iPad off Happy to answer questions in middle of talk except…
- For example, geographical maps have street data…buildings data…vegetation data…all resulting in an integrated map
- Semi-automated genomic annotation begins with pattern discovery from multiple genomic data sets and results in: A simple annotation with a single label for each part of the genome using these patterns We can use this annotation to visualise a huge number of datatracks, eg viewing the resulting patterns found and the annotation instead of looking at each datatrack individually We can interpret the context and potential impact of the results, for example the meaning of the patterns and annotation we found. Some questions that semi-automated annotation can answer include: Pattern discovery: What signals from multiple experiments do we see over and over again? Annotation: What does a particular piece of the genome do, in a nutshell? Visualization: How can we make complex data comprehensible visually? Interpretation: What is the context and potential impact of the results we are finding?
- To perform genomic segmentation, Segway first ‘observes’ multiple datasets of genomic data.
- Then, it splits the datasets up into non-overlapping segments.
- To each segment, Segway assigns a label from a finite set
- Segway then maximizes the similarity between segments of the same label by pushing around the boundaries of the segments. These are our ‘learned patterns’
- For example, here we have a 0-label with a low-high-low pattern
- 1-label with a high-low-high pattern
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- A hidden Markov multimodel: “no longer a HMM, not yet a DBN” dashed line = switching relationship The default Segway model for three observations.
- DBNs can be thought of as a generalization of HMMs. This particular DBN is just another representation of an HMM. The DBN diagram itself does not contain the state transition information in an HMM diagram, but it is contained in the transition conditional probability table.
- A hidden Markov multimodel: “no longer a HMM, not yet a DBN” dashed line = switching relationship The default Segway model for three observations.
- A hidden Markov multimodel: “no longer a HMM, not yet a DBN” dashed line = switching relationship The default Segway model for three observations.
- A hidden Markov multimodel: “no longer a HMM, not yet a DBN” dashed line = switching relationship The default Segway model for three observations.
- A hidden Markov multimodel: “no longer a HMM, not yet a DBN” dashed line = switching relationship The default Segway model for three observations.
- A hidden Markov multimodel: “no longer a HMM, not yet a DBN” dashed line = switching relationship The default Segway model for three observations.
- A hidden Markov multimodel: “no longer a HMM, not yet a DBN” dashed line = switching relationship The default Segway model for three observations.
- light blue fill = semiobserved random variable (observed random variable for the data switched by a discrete observed random variable for the nonmissingness of the data)
- light blue fill = semiobserved random variable (observed random variable for the data switched by a discrete observed random variable for the nonmissingness of the data) Discuss mixture model
- light blue fill = semiobserved random variable (observed random variable for the data switched by a discrete observed random variable for the nonmissingness of the data)
- light blue fill = semiobserved random variable (observed random variable for the data switched by a discrete observed random variable for the nonmissingness of the data)
- And finally, I'd like to thank you for your very kind attention.