1、BIG Biomedicine and the Foundations of BIG Data Analysis,Michael W. MahoneyICSI and Dept of Statistics, UC BerkeleyMay 2014(For more info, see: http:/www.stat.berkeley.edu/mmahoney),Insiders vs outsiders views (1 of 2),Ques: Genetics vs molecular biology vs biochemistry vs biophysics: Whats the diff
2、erence?,Insiders vs outsiders views (1 of 2),Ques: Genetics vs molecular biology vs biochemistry vs biophysics: Whats the difference?Answer: Not much, (if you are a “methods” person*)they are all biologyyou get data from any of those areas, ignoring important domain details, and evaluate your method
3、 qua methodyour reviewers evaluate the methods and dont care about the science.*E.g., one who self-identifies as doing data analysis or machine learning or statistics or theory of algorithms or artificial intelligence or .,Insiders vs outsiders views (2 of 2),Ques: Data analysis vs machine learning
4、vs statistics vs theory of algorithms vs artificial intelligence (vs scientific computing vs computational mathematics vs databases .): Whats the difference?,Insiders vs outsiders views (2 of 2),Ques: Data analysis vs machine learning vs statistics vs theory of algorithms vs artificial intelligence
5、(vs scientific computing vs computational mathematics vs databases .): Whats the difference?Answer: Not much, (if you are a “science” person*)they are all just toolsyou get a tool from any of those areas and bury details in a methods sectionyour reviewers evaluate the science and dont care about the
6、 methods.*E.g., one who self identifies as doing genetics or molecular biology or biochemistry or biophysics or .,BIG data? MASSIVE data?,NYT, Feb 11, 2012: “The Age of Big Data” “What is Big Data? A meme and a marketing term, for sure, but also shorthand for advancing trends in technology that open
7、 the door to a new approach to understanding the world and making decisions. ” Why are big data big? Generate data at different places/times and different resolutionsFactor of 10 more data is not just more data, but different data,Thinking about large-scale data,Data generation is modern version of
8、microscope/telescope: See things couldnt see before: e.g., fine-scale movement of people, fine-scale clicks and interests; fine-scale tracking of packages; fine-scale measurements of temperature, chemicals, etc.Those inventions ushered new scientific eras and new understanding of the world and new t
9、echnologies to do stuffEasy things become hard and hard things become easy: Easier to see the other side of universe than bottom of oceanMeans, sums, medians, correlations is easy with small data,Our ability to generate data far exceeds our ability to extract insight from data.,How do we view BIG da
10、ta?,Algorithmic vs. Statistical Perspectives,Computer Scientists Data: are a record of everything that happened. Goal: process the data to find interesting patterns and associations.Methodology: Develop approximation algorithms under different models of data access since the goal is typically comput
11、ationally hard.Statisticians (and Natural Scientists)Data: are a particular random instantiation of an underlying process describing unobserved patterns in the world.Goal: is to extract information about the world from noisy data.Methodology: Make inferences (perhaps about unseen events) by positing
12、 a model that describes the random variability of the data around the deterministic model.,Lambert (2000), Mahoney (2010),Single Nucleotide Polymorphisms: the most common type of genetic variation in the genome across different individuals.They are known locations at the human genome where two alter
13、nate nucleotide bases (alleles) are observed (out of A, C, G, T).,SNPs,individuals, AG CT GT GG CT CC CC CC CC AG AG AG AG AG AA CT AA GG GG CC GG AG CG AC CC AA CC AA GG TT AG CT CG CG CG AT CT CT AG CT AG GG GT GA AG GG TT TT GG TT CC CC CC CC GG AA AG AG AG AA CT AA GG GG CC GG AA GG AA CC AA CC
14、AA GG TT AA TT GG GG GG TT TT CC GG TT GG GG TT GG AA GG TT TT GG TT CC CC CC CC GG AA AG AG AA AG CT AA GG GG CC AG AG CG AC CC AA CC AA GG TT AG CT CG CG CG AT CT CT AG CT AG GG GT GA AG GG TT TT GG TT CC CC CC CC GG AA AG AG AG AA CC GG AA CC CC AG GG CC AC CC AA CG AA GG TT AG CT CG CG CG AT CT
15、CT AG CT AG GT GT GA AG GG TT TT GG TT CC CC CC CC GG AA GG GG GG AA CT AA GG GG CT GG AA CC AC CG AA CC AA GG TT GG CC CG CG CG AT CT CT AG CT AG GG TT GG AA GG TT TT GG TT CC CC CG CC AG AG AG AG AG AA CT AA GG GG CT GG AG CC CC CG AA CC AA GT TT AG CT CG CG CG AT CT CT AG CT AG GG TT GG AA GG TT
16、TT GG TT CC CC CC CC GG AA AG AG AG AA TT AA GG GG CC AG AG CG AA CC AA CG AA GG TT AA TT GG GG GG TT TT CC GG TT GG GT TT GG AA ,Matrices including thousands of individuals and hundreds of thousands if SNPs are available, and more/bigger/better are coming soon.This can be written as a “matrix,” ass
17、ume its been preprocessed properly, so lets call black box matrix algorithms.,Applications in: Human Genetics,Africa,Middle East,S C Asia & Gujarati,Europe,Oceania,East Asia,America,Not altogether satisfactory: the principal components are linear combinations of all SNPs, and of course can not be as
18、sayed! Can we find actual SNPs that capture the information in the singular vectors? Formally: spanning the same subspace, optimizing variance, computationally efficient.,Mexicans,Paschou, et al. (2010) J Med Genet,Apply PCA/SVD:,Issues with eigen-analysis,Computing large SVDs: computational timeIn
19、commodity hardware (e.g., a 4GB RAM, dual-core laptop), using MatLab 7.0 (R14), the computation of the SVD of the dense 2,240-by-447,143 matrix A takes about 20 minutes.Computing this SVD is not a one-liner, since we can not load the whole matrix in RAM (runs out-of-memory in MatLab). Instead, compu
20、te the SVD of AAT.In a similar” experiment,” compute 1,200 SVDs on matrices of dimensions (approx.) 1,200-by-450,000 (roughly, a full leave-one-out cross-validation experiment).Selecting actual columns that “capture the structure” of the top PCsCombinatorial optimization problem; hard even for small
21、 matrices. Often called the Column Subset Selection Problem (CSSP).Not clear that such “good” columns even exist.Avoid “reification” problem of “interpreting” singular vectors!,CUR matrix decompositions,Goal. Solve the following problem: “While very efficient basis vectors, the (singular) vectors th
22、emselves are completely artificial and do not correspond to actual (DNA expression) profiles. . . . Thus, it would be interesting to try to find basis vectors for all experiment vectors, using actual experiment vectors and not artificial bases that offer little insight.” Kuruvilla et al. (2002)Theor
23、em: Given an arbitrary matrix, call a black box that I wont describe. You get a small number of actual columns/rows that are only marginally worse than the truncated PCA/SVD. The black box runs faster than computing a truncated PCA/SVD for arbitrary input. Its very robust to heuristic modifications.
24、 Corollary: We can use the same methods to approximate the PCA/SVD.,Mahoney and Drineas “CUR Matrix Decompositions for Improved Data Analysis” (PNAS, 2009),SNPs by chromosomal order,PCA-scores,* top 30 PCA-correlated SNPs,Africa,Europe,Asia,America,Selecting PCA SNPs for individual assignment to fou
25、r continents (Africa, Europe, Asia, America),Mahoney and Drineas (2009) PNAS Paschou et al (2007; 2008) PLoS Genetics Paschou et al (2010) J Med Genet Drineas et al (2010) PLoS One Javed et al (2011) Annals Hum Genet,Data analysis and machine learning and statistics and theory of algorithms and scie
26、ntific computing . and genetics and astronomy and mass spectrometry and . likes this-but each for different reasons! Good “hydrogen atom” for methods development!,Bioinformatics: a cautionary tale?,How did/does bioinformatics relate to computer science, statistics, and applied mathematics, “technica
27、lly” and “sociologically”?How did NIH choose to fund graduate students and postdocs in the budget expansion of the 90s?What effect did this have on the number of American/foreign going into biomedical research?How will the pay structure of biomedical researchers effect which cs/stats “data scientist
28、s” engage you in your efforts?What effect does med schools deciding not to do joint faculty hires with cs departments have on bioinformatics and big biomedical data? How is this Big Biomedical Data phenomenon similar to and different than the Bioinformatics experience?,Big changes in the past . and
29、future,Consider the creation of: Modern PhysicsComputer ScienceMolecular Biology These were driven by new measurement techniques and technological advances, but they led to: big new (academic and applied) questionsnew perspectives on the worldlots of downstream applications We are in the middle of a
30、 similarly big shift!,OR and Management Science Transistors and MicroelectronicsBiotechnology,MMDS Workshop on “Algorithms for Modern Massive Data Sets” (http:/mmds-data.org),at UC Berkeley, June 17-20, 2014Objectives:Address algorithmic, statistical, and mathematical challenges in modern statistica
31、l data analysis.Explore novel techniques for modeling and analyzing massive, high-dimensional, and nonlinearly-structured data. - Bring together computer scientists, statisticians, mathematicians, and data analysis practitioners to promote cross-fertilization of ideas.Organizers: M. W. Mahoney, A. Shkolnik, P. Drineas, R. Zadeh, and F. PerezRegistration is available now!,
