Charalampos (Babis) E. Tsourakakis Joint work with Gary Miller, Richard Peng, Russell Schwartz, Stanley Shackney, Dave Tolliver, Maria A. Tsiarli Algorithms.

Slides:



Advertisements
Similar presentations
Principal Component Analysis Based on L1-Norm Maximization Nojun Kwak IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008.
Advertisements

Unsupervised Learning
Fast Algorithms For Hierarchical Range Histogram Constructions
CSCI 347 / CS 4206: Data Mining Module 07: Implementations Topic 03: Linear Models.
EE462 MLCV Lecture Introduction of Graphical Models Markov Random Fields Segmentation Tae-Kyun Kim 1.
1 Learning Dynamic Models from Unsequenced Data Jeff Schneider School of Computer Science Carnegie Mellon University joint work with Tzu-Kuo Huang, Le.
Hidden Markov Model based 2D Shape Classification Ninad Thakoor 1 and Jean Gao 2 1 Electrical Engineering, University of Texas at Arlington, TX-76013,
Methods for copy number variation: hidden Markov model and change- point models.
Presenter: Yufan Liu November 17th,
Visual Recognition Tutorial
Tumour karyotype Spectral karyotyping showing chromosomal aberrations in cancer cell lines.
Yanxin Shi 1, Fan Guo 1, Wei Wu 2, Eric P. Xing 1 GIMscan: A New Statistical Method for Analyzing Whole-Genome Array CGH Data RECOMB 2007 Presentation.
Content Based Image Clustering and Image Retrieval Using Multiple Instance Learning Using Multiple Instance Learning Xin Chen Advisor: Chengcui Zhang Department.
Sam Pfister, Stergios Roumeliotis, Joel Burdick
Paper Discussion: “Simultaneous Localization and Environmental Mapping with a Sensor Network”, Marinakis et. al. ICRA 2011.
Clustering short time series gene expression data Jason Ernst, Gerard J. Nau and Ziv Bar-Joseph BIOINFORMATICS, vol
Algorithms for Smoothing Array CGH data
Lecture 5: Learning models using EM
Charalampos (Babis) E. Tsourakakis Machine Learning Seminar 10 th January ‘11 Machine Learning Lunch Seminar1.
Preference Analysis Joachim Giesen and Eva Schuberth May 24, 2006.
Collaborative Ordinal Regression Shipeng Yu Joint work with Kai Yu, Volker Tresp and Hans-Peter Kriegel University of Munich, Germany Siemens Corporate.
Amos Storkey, School of Informatics. Density Traversal Clustering and Generative Kernels a generative framework for spectral clustering Amos Storkey, Tom.
Review Rong Jin. Comparison of Different Classification Models  The goal of all classifiers Predicating class label y for an input x Estimate p(y|x)
1 Dot Plots For Time Series Analysis Dragomir Yankov, Eamonn Keogh, Stefano Lonardi Dept. of Computer Science & Eng. University of California Riverside.
CSC2535: 2013 Advanced Machine Learning Lecture 3a: The Origin of Variational Bayes Geoffrey Hinton.
The Paradigm of Econometrics Based on Greene’s Note 1.
Graph-based consensus clustering for class discovery from gene expression data Zhiwen Yum, Hau-San Wong and Hongqiang Wang Bioinformatics, 2007.
Binary Variables (1) Coin flipping: heads=1, tails=0 Bernoulli Distribution.
Muhammad Moeen YaqoobPage 1 Moment-Matching Trackers for Difficult Targets Muhammad Moeen Yaqoob Supervisor: Professor Richard Vinter.
Charalampos (Babis) E. Tsourakakis SODA th January ‘11 SODA '111.
Genetic Regulatory Network Inference Russell Schwartz Department of Biological Sciences Carnegie Mellon University.
Charalampos (Babis) E. Tsourakakis WABI 2013, France WABI '131.
Bayesian networks Classification, segmentation, time series prediction and more. Website: Twitter:
CSCI 256 Data Structures and Algorithm Analysis Lecture 14 Some slides by Kevin Wayne copyright 2005, Pearson Addison Wesley all rights reserved, and some.
CS 782 – Machine Learning Lecture 4 Linear Models for Classification  Probabilistic generative models  Probabilistic discriminative models.
The Group Lasso for Logistic Regression Lukas Meier, Sara van de Geer and Peter Bühlmann Presenter: Lu Ren ECE Dept., Duke University Sept. 19, 2008.
The Dirichlet Labeling Process for Functional Data Analysis XuanLong Nguyen & Alan E. Gelfand Duke University Machine Learning Group Presented by Lu Ren.
Randomized Algorithms for Bayesian Hierarchical Clustering
ECE 8443 – Pattern Recognition ECE 8527 – Introduction to Machine Learning and Pattern Recognition Objectives: Evaluation Decoding Dynamic Programming.
INTRODUCTION TO Machine Learning 3rd Edition
1  The Problem: Consider a two class task with ω 1, ω 2   LINEAR CLASSIFIERS.
De novo discovery of mutated driver pathways in cancer Discussion leader: Matthew Bernstein Scribe: Kun-Chieh Wang Computational Network Biology BMI 826/Computer.
1 Neighboring Feature Clustering Author: Z. Wang, W. Zheng, Y. Wang, J. Ford, F. Makedon, J. Pearlman Presenter: Prof. Fillia Makedon Dartmouth College.
ETHEM ALPAYDIN © The MIT Press, Lecture Slides for.
Fast Query-Optimized Kernel Machine Classification Via Incremental Approximate Nearest Support Vectors by Dennis DeCoste and Dominic Mazzoni International.
Dimensionality Reduction in Unsupervised Learning of Conditional Gaussian Networks Authors: Pegna, J.M., Lozano, J.A., Larragnaga, P., and Inza, I. In.
Machine Learning 5. Parametric Methods.
CGH Data BIOS Chromosome Re-arrangements.
Javad Azimi, Ali Jalali, Xiaoli Fern Oregon State University University of Texas at Austin In NIPS 2011, Workshop in Bayesian optimization, experimental.
1 Chapter 8: Model Inference and Averaging Presented by Hui Fang.
The Unscented Particle Filter 2000/09/29 이 시은. Introduction Filtering –estimate the states(parameters or hidden variable) as a set of observations becomes.
An Efficient Algorithm for a Class of Fused Lasso Problems Jun Liu, Lei Yuan, and Jieping Ye Computer Science and Engineering The Biodesign Institute Arizona.
Density Estimation in R Ha Le and Nikolaos Sarafianos COSC 7362 – Advanced Machine Learning Professor: Dr. Christoph F. Eick 1.
ETHEM ALPAYDIN © The MIT Press, Lecture Slides for.
CMPS 142/242 Review Section Fall 2011 Adapted from Lecture Slides.
Markov Chain Monte Carlo in R
Tijl De Bie John Shawe-Taylor ECS, ISIS, University of Southampton
Chapter 7. Classification and Prediction
Learning Deep Generative Models by Ruslan Salakhutdinov
LINEAR CLASSIFIERS The Problem: Consider a two class task with ω1, ω2.
Bayesian Generalized Product Partition Model
LECTURE 15: HMMS – EVALUATION AND DECODING
Overview of Supervised Learning
Statistical Learning Dong Liu Dept. EEIS, USTC.
Hidden Markov Models Part 2: Algorithms
دانشگاه صنعتی امیرکبیر Instructor : Saeed Shiry
LECTURE 14: HMMS – EVALUATION AND DECODING
Biointelligence Laboratory, Seoul National University
Finding Periodic Discrete Events in Noisy Streams
Introduction to Object Tracking
Presentation transcript:

Charalampos (Babis) E. Tsourakakis Joint work with Gary Miller, Richard Peng, Russell Schwartz, Stanley Shackney, Dave Tolliver, Maria A. Tsiarli Algorithms for Denoising aCGH Data1 Speaking Skills Machine Learning Journal Club 23 Feb. 2010

 Motivation & Problem Definition  Related Work  Our Problem Formulation and a O(n 2 ) solution  Experimental Results  Theoretical Ramifications: a O(n 1.5 ) algorithm within additive ε error)  Conclusion & Future Work Algorithms for Denoising aCGH Data2 ~

3 Test DNA: Patient Reference DNA: Healthy subject For each probe we obtain a noisy measurement of log(T/R) where T: true DNA copy number R=2 for humans (diploid organisms)

 Ideal Scenario  In practice, for a variety of reasons (e.g., sample impurity, measurement noise) we obtain a noisy measurement log(T/R) per probe. Algorithms for Denoising aCGH Data4 Copy Numberlog(T/R) 0-Inf 1 20 (Healthy probe) T: true DNA copy number R=2 for humans (diploid organisms)

 Input A vector (t 1,t 2,…,t n ), where t i is the measurement at the i-th proble  Output A vector (t 1,t 2,…,t n ) with discrete values corresponding to the true DNA copy number Algorithms for Denoising aCGH Data5 ~~~~

6 Probe id log(T/R) Blue x : noisy measurements (input) Red □ : true value (output)

Algorithms for Denoising aCGH Data7 Probe id log(T/R) Nearby probes tend to have the same DNA copy number! 2) Fit Piecewise Constant Segments 1)Treat Data as 1D time series Well studied problem with many Applications

 Motivation & Problem Definition  Related Work  Our Problem Formulation and a O(n 2 ) solution  Experimental Results  Theoretical Ramifications: a O(n 1.5 ) algorithm within additive ε error)  Conclusion & Future Work Algorithms for Denoising aCGH Data8 ~

 Lasso (Tibshirani et al., Huang et al.)  Kalman Filters (Xing et al.)  Hidden Markov Models (Fridlyand et al.)  Bayesian Hidden Markov Models (Guha et al.)  Wavelets (Hsu et al.)  Hierarchical clustering (Tibshirani et al.)  Circular Binary Segmentation (Olshen et al.)  Statistical likelihood tests  Loweless, i.e., Locally weighted regression  Genetic Local search (Jong et al.) Algorithms for Denoising aCGH Data9 Not Exhaustive

 Gaussian Mixtures fitting using EM (Hodgson et al.)  Variable-bandwidth kernel methods (Muller et al.)  Variable-knot splines (Stone et al.)  Fused quantile regression (Wang et al.)  Non parametric regression  Thresholding Algorithms for Denoising aCGH Data10 Not Exhaustive

Algorithms for Denoising aCGH Data11  CBS: Matlab Toolbox, modification of Binary Segmentation.  CGHSEG: Gaussianity, AIC&BIC, DP  Both methods perform consistently better than others on real data (Lai et al., Willenbrock et al.)

 Motivation & Problem Definition  Related Work  Our Problem Formulation and a O(n 2 ) solution  Experimental Results  Theoretical Ramifications: a O(n 1.5 ) algorithm within additive ε error)  Conclusion & Future Work Algorithms for Denoising aCGH Data12 ~

Algorithms for Denoising aCGH Data13 Breakpoint Squared error Penalty per segment For the vector (p 1,…,p n ) we define the following recurrence equation: Tradeoff(λ)

Algorithms for Denoising aCGH Data14 Keep the first and second order moments in an “online” way. Run time O(n 2 )

Algorithms for Denoising aCGH Data15 λ=0.2 Train on synthetic data generated by a realistic simulator Willenbrock et al λ=0.2 results in Precision=0.98 Recall=0.91

 Motivation & Problem Definition  Related Work  Our Problem Formulation and a O(n 2 ) solution  Experimental Results  Theoretical Ramifications: a O(n 1.5 ) algorithm within additive ε error)  Conclusion & Future Work Algorithms for Denoising aCGH Data16 ~

Algorithms for Denoising aCGH Data17 A)CSB B)CGHTrimmer C)CGHSEG Lai et al. (2005) Dataset Available from

Algorithms for Denoising aCGH Data18 Snijders et al., 15 Cell lines Golden Standard Dataset with two main characteristics a)Knowledge of ground truth after tedious biological tests b)“Easy” dataset (low noise levels) CGHTrimmer performs worse that at least one competitor CGHTrimmer performs equally well with both competitors CGHTrimmer performs better than one competitor CGHTrimmer performs better than both competitors #Cell Lines Color Description

Algorithms for Denoising aCGH Data19 A)CBS B)CGHTrimmer C)CGHSEG Breast Cancer Cell Line BT474 Chromosome 1

Algorithms for Denoising aCGH Data20 A)CBS B)CGHTrimmer C)CGHSEG Breast Cancer Cell Line BT474 Chromosome 17 Results supported by oncology literature

Algorithms for Denoising aCGH Data21 A)CBS B)CGHTrimmer C)CGHSEG Breast Cancer Cell Line T47D Chromosome 1

CGHTrimmerCGHSEGCBS Coriell5.78sec8.15min47.7min Breast Cancer min4.95hours Algorithms for Denoising aCGH Data22 A) 1 to 3 orders of magnitude faster. REMARKS B) Reason for speedup: different approach compared to competitors (lack of statistical assumptions, tests, likelihood functions but an intuitive formulation and a simple dynamic programming algorithm)

 Motivation & Problem Definition  Related Work  Our Problem Formulation and a O(n 2 ) solution  Experimental Results  Theoretical Ramifications: a O(n 1.5 ) algorithm within additive ε error)  Conclusion & Future Work Algorithms for Denoising aCGH Data23 ~

 Is O(n 2 ) tight? Probably not…  Lemma If |p i -p j | > 2 then in the optimal solution points i,j belong to different segments.  Question: Can we use one of the existing “tricks” to speed up our dynamic program? Algorithms for Denoising aCGH Data24

 Unfortunately, existing “tricks” for dynamic programming do not work for us (e.g., Monge property)  But we can find an good approximation algorithm! Algorithms for Denoising aCGH Data25

Algorithms for Denoising aCGH Data26 Define a shifted by a constant objective function, i.e., DP i =OPT i - Claim: DP i satisfies the following optimization formula where S i =p 1 +…+p i

Algorithms for Denoising aCGH Data27 Find the maximum and the minimum value that the shifted objected DP i can take. Claim: DP i takes values in [0,U 2 n] where

Algorithms for Denoising aCGH Data28 Perform binary search on by guessing for each index i a DP i ~ 0 U2nU2n () DP i ~ -

Algorithms for Denoising aCGH Data29 constant Dot product of two 4D points (x,i,S i,-1) and (j,DP j,2S j,S j 2 +jDPj) ~~ Reporting points in Halfspaces, Matousek FOCS 1992

Algorithms for Denoising aCGH Data30 0U2nU2n () DP i ~ - Remember: We want DP i not DP i ε/n by performing log( U 2 n/(ε/n) ) iterations Set i=n, our algorithm is an ε-additive approximation algorithm. Run Time:

 Motivation & Problem Definition  Related Work  Our Problem Formulation and a O(n 2 ) solution  Experimental Results  Theoretical Ramifications: a O(n 1.5 ) algorithm within additive ε error)  Conclusion & Future Work Algorithms for Denoising aCGH Data31 ~

 CHGtrimmer method: Simple, intuitive dynamic programming algorithm outperforms state-of-the-art competitors:  Important biological biomarkers for aCGH data.  Good precision-recall results.  Significantly faster.  New paradigm for Dynamic Programming by reducing the problem to a computational geometry halfspace query problem. Algorithms for Denoising aCGH Data32

 Structure: O(n 2 ) unlikely to be tight.  Extend our approximation to 2D recurrences (benefit many applications)  Preprocess breast cancer data using Trimmer to get a discretized version, essential to perform standard tumor phylogenetics  Make the DS practical. Algorithms for Denoising aCGH Data33

 CGHTrimmer: Discretizing Noisy Array CGH Data C.E.T, D. Tolliver, M. A. Tsiarli, S. Shackney, R. Schwartz  Algorithms for Denoising aCGH Data G.L. Miller, R. Peng. R. Schwartz, C.E.T Code will be made available at Algorithms for Denoising aCGH Data34