#!" $"" % &'' % " ( Data analysis process Bayesian decision theoretic foundation Causal inference Open access data, publication, model, computation

Transcription

1

2 !" #!" $"" % # &'' % " ( Data analysis process Bayesian decision theoretic foundation Causal inference Open access data, publication, model, computation

3 Langley, P. (1978). Bacon: A general discovery system. Proceedings of the Second Biennial Conference of the Canadian Society for Computational Studies of Intelligence (pp ). Toronto, Ontario. Chrisman, L., Langley, P., & Bay, S. (2003). Incorporating biological knowledge into evaluation of causal regulatory hypotheses. Proceedings of the Pacific Symposium on Biocomputing (pp ). Lihue, Hawaii. (Gene prioritization ) R.D.King et al.: The Automation of Science, Science, 2009

4 *+, -" ) "! " " ''' The line between the virtual world of computing and our physical, organic world is blurring. E.Dumbill: Making sense of big data, Big Data, vol.1, no.1, 2013 ) &/ (.

5 +"- M. Cox and D. Ellsworth, Managing Big Data for Scientific Visualization, Proc. ACM Siggraph, ACM, 1997 The 3xV: volume, variety, and velocity (2001). The 8xV: Vast, Volumes of Vigorously, Verified, Vexingly Variable Verbose yet Valuable Visualized high Velocity Data (2013) Not conventional data: Big data is data that exceeds the processing capacity of conventional database systems. The data is too big, moves too fast, or doesn t fit the strictures of your database architectures. To gain value from this data, you must choose an alternative way to process it (E.Dumbill: Making sense of big data, Big Data, vol.1, no.1, 2013) 0

6 +%""-11 Hype Application Boom Criticism, re-evaluation Technology trigger Lack of results, problems, $1 2

7 +"- / 3

8 +"- /.. [data] is often big in relation to the phenomenon that we are trying to record and understand. So, if we are only looking at 64,000 data points, but that represents the totality or the universe of observations. That is what qualifies as big data. You do not have to have a hypothesis in advance before you collect your data. You have collected all there is all the data there is about a phenomenon. 4

9 % 5" 2010<: Clinical phenotypic assay /drugome: open clinical trials, adverse drug reaction DBs, adaptive licensing, Large/scale cohort studies (~100,000 samples) Environment&life style Phenome (disease, side effect) Metabolome Proteome Transcriptome Genome(s), epigenome, microbiome Moore s law Drugs Carlson s law

10 The hypothesis-free pitfall in omics 6 7! 6 7! 6 5 " GO STRING.. 89

11 !! 88

12 *! + -' :9 $ ':8 %' *7 ;* * 55 ' ' 5! 5 "5 5:99< ( %*7 ;= ; >:>8:<3!8:<4:99<

13 +?!!!! '* B'-7 7 ""? #7, %, & ; ( C! * DD

14 +!- $ 8<!!" # 8<:9! #' E " " F 5! "" ; "" "" " % "" " % G

15 * 1 1 * ' ' ' * " ' ) " ' " + -'

16 H" I /= ' '7 "" ' )J '' )$ 'H'$= K'K1" ;"L %'' %+ " - (

17 ) $""1 +$""M F N","" % &=5, ''''

18 "" "" ; C C C C "" " C C JC 5"L C C lim N N A = lim p ˆ ( A ) N N = N *+- O """ $ " I " p ( A )? p ( A ξ ) Note: independence and convergence of frequencies are empirical observations (i.e., laws of large numbers consequences of independencies)

19 @838>A,L *,L N "%! " & """ "# + "" """" A*;$", * % $# %+ "" "" " " # ( ;' E /!8<09"% H')N'? ' A+ ""- 7 +) -Q% 7

20 R#PJ "L -%') R*""+7 N 8<09!+ ' +J- 'N'( %'I' 7 HH"; ; *'$' = ;'#= G&,&, S( $( $O G

21 %! ()*+,()-(. %"" %/ % % % % % % %+ - ''' p( Model Data) p(data Model) p(model) a * = arg max U ( o ) p ( o a ) p( prediction data) = i p( pred. i j = i j Model ) p( Model j i i data)

22 %$" Number of papers in Pubmed year Number of papers in Pubmed for query Bayesian What is the reason? Yet another statistical approach (in the hype phase)? A (statistical) paradigm shift? A new scientific paradigm? Something practical?

23 IJ""'

24 Recall: utility theory! $""'

25 / 0 #/ ' p ( X Y ) = p( Y X ) p( X p( Y ) ) A scientific research paradigm p( Model Data) p( Data Model) p( Model) A practical method for inverting causal knowledge to diagnostic tool. p ( Cause Effect) p( Effect Cause) p( Cause)

26 )J% In the frequentist approach: Model identification (selection) is necessary p ( prediction data) = p( prediction BestModel( data)) In the Bayesian approach models are weighted p( prediction data) = i p( pred. Model i ) p( Model i data) Note: in the Bayesian approach there is no need for model selection

27 % Russel&Norvig: Artificial intelligence, ch.20

28 %& Russel&Norvig: Artificial intelligence

29 # Russel&Norvig: Artificial intelligence

30 # 5 Russel&Norvig: Artificial intelligence

31 ) ""

32 H

33 O&

34 O Russel&Norvig: Artificial intelligence, ch.16

35 O Russel&Norvig: Artificial intelligence, ch.16

36 ""F I $"" O 5 & O $ &O % a o i j p( o j ai ) U o j a ) ( i i = EU ( a ) U ( o a ) p( o a ) j a * = arg max EU ( a i i j i j i ) Actions a i Outcomes Probabilities Utilities, costs Expected utilities P(o j a i ) U(o j ), C(a i ) EU(a i ) = P(o j a i )U(o j ) a i o j

37 * * )55? *55T5)$ 7 9 U7 9 ; 7 9 U7 9 M8!T ; T!+""" - *5 5V5)?7 9 U 7 9 $ 7 9 U 7 9 M8!V

38 )J%! " # $$!%% # &'$! ) ( )&'' *+% #&! #, " #, $$ -&.'& )!'$0 )0200 %0$ #&! -$' /0$

39 O ' O *%;; H'; "1 "" "1 B>.:1%','!:39%',' ' * I B=8:40! 8>.< I ' I B= '* " ' " "" ""'

40 O ' O " "J " " W " & F8U8'''M& 8'''F85& 8''' WF8U8'''MW8'''F85W8''' W!""8' 7 & "W ' & '#$'& %'K' ';? X:8<<3>:994

41 M Data Prior IDA Optimal action (reporting, publishing,.., surgery). Inference about the world/domain.

42 *& # ''!! ;! 8 ;

43 * I"5 O 1? 1, 5, ;"5H5,)( 1 ; =

44 * " 5 )1 ;! #, $, ) "

45 * $" $Y UI"I" $Y UI", $Y UI", % $ ; * *

46 Clusters, modules Conditional independencies (Causal) Mechanisms Parameters

47 % L " ; " ZR R " [ U$[ " """,$*" "

48 B "N ""& B '; B" J'"1 K"!? R R " J * & * N

49 '

50 ,,$*% %" " J " # " M L$% " J&'( : " '''&'" )L" )"8F8F.F:F:M89 "': 0!8M>8

51 ; *L" " ' $ *+* )#, #$ ' -. ' )) n '''/ ) #'/- )' - )' - * )' )' )

52 , % 8'," $ G :') M8 $ *+* %$ ' -. ' ))#' - $ *000 %$ ) * ' $ *+* ) n #, #$ ' n - $ *+* %$ ) 11 #, #$ ' -. ' )) 11"

53 " J" " "L"' * L"J" 1 " 2 #""3M4! 1 S 2 *J L" " 1U2MMT12MMT\ 12M3M * L" 12 3 " I"" 8' E! &' ) :' ; &' )L" >' 7 "&' )1

54 , +$""M F - $ [CXU][XU] $ [X] $[CXU=M$XU=$[U== $=^9' $ [CXU] $[U]XM$[U]= $=^9' I $ [CXU]MM_ $ [CXU], ='

55 ?%?%? Decomposition of the joint: P(Y,X 1,..,X n ) = P(Y) i P(X i, Y, X 1,..,X i-1 ) //by the chain rule = P(Y) i P(X i, Y) // by the N-BN assumption 2n+1 parameteres! Diagnostic inference: P(Y x i1,..,x ik ) = P(Y) j P(x ij, Y) / P(x i1,..,x ik ) If Y is binary, then the odds P(Y=1 x i1,..,x ik ) / P(Y=0 x i1,..,x ik ) = P(Y=1)/P(Y=0) j P(x ij, Y=1) / P(x ij, Y=0) Flu Fever Coughing p( Flu = present Fever = absent, Coughing = present) p( Flu = present) p( Fever = absent Flu = present) p( Coughing = present Flu = present)

56

57 ?!IH 88545:98> '' 03

58 absent Bleeding weak strong Onset Regularity P(D Bleeding=strong) Onset=early Onset=late regular P(D a,e) Mutation P(D w,r) h.wild mutated P(D a,l,h.w.) P(D a,l,m) irregular Mutation h.wild mutated P(D w,i,h.w.) P(D w,i,m) Decision tree: Each internal node represent a (univariate) test, the leafs contains the conditional probabilities given the values along the path. Decision graph: If conditions are equivalent, then subtrees can be merged. E.g. If (Bleeding=absent,Onset=late) ~ (Bleeding=weak,Regularity=irreg)

59 * / inference\data Observational Interventional Observational OK OK Interventional???? OK Counterfactual???????? Automated, tabula rasa causal inference from (passive) observation is possible, i.e. hidden, confounding variables can be excluded? E X Y??? * Plato s two surprises: 1. Not all true theorems can be proved 2. Causal inference is possible from observations

60 , 5 K? N''E '%8<22 *! E 5" D

61 #5 Learning step? Learning process Data Domain knowledge Learning algorithm Result Interpretation and evaluation Running the learning algorithm Selection of learning method Feature selection Data engineering Integration Understanding the domain

62 * ;

63 & $ %! " # $ $ % ) * % ' +, ( - ' ' " " " " ).. /. " $

64 ) J A p(d = 1 A = 0) p(d = 1 A = 2) Real difference Estimation errors D p ˆ(D = 1 A = 0, D N ) p ˆ(D = 1 A = 2, D N ) Estimated difference Relative frequencies: Law of large numbers: ˆ (D = 1 A = 0, DN ) = pˆ(d = 1, A = 0, DN )/ˆ(A p = 0, DN ) = N D= 1, A= 0 / N A= 0 p p (D = 1 A = 0) N D = 1, A= 0 / N A= 0 Estimation error because of finite data D N : pˆ (D = 1 A = 0, D ) - p(d = 1 A = 0) N Central limit th. (asymptotic), Inequalities for finite(!) data ( accuracy, confidence) sample complexity: N, p DN : ε < pˆ(d = 1 A, DN ) (D = 1 A) ) ( < ε δ ε, δ, p δ

65 # Interpretation 17% Integration 9% Data engineering 26% Running 1% Selection of learning method 26% Feature selection 17%

66 )1 Domain Iteration in learning Integration Evaluation Interpretion Running Understanding the task Data engineering Feature selection The Learning Phases (in waterfall model) Data Selection of learning method The Learning Step Domain knowledge Integration Interpretation and evaluation Running the learning algorithm Selection of learning method Learning algorithm Feature selection Result Understanding the domain Data engineering

67 *" Q ) '';?$ & "

68 % $ 8<!!" # 8<:9! #' E " " F %! "" ; "" "" " % "" " % G

69 ; " %5 % & '( R!"E "$ " R *8<<< & '( #66? :998 $'%RE " "" " L1R. *:990 & '( R; " R2 *:993' & '; R$"1* B R *:993' & ';R& " R3 4*:994' '; R; " L"R4.*:99< 2<

70 " " ( I " %! %! ( ;= H. Colhoun: "Problems of reporting genetic associations with complex outcomes," Lancet, J. Attia et al: "How to Use an Article About Genetic Association A: Background Concepts,", B: Are the Results of the Study Valid?," C: What Are the Results and Will They Help Me in Caring for My Patients?," Jama, J. Little et al: "STrengthening the REporting of Genetic Association studies (STREGA) " European Journal of Clinical Investigation, J. Huang et al: "Minimum Information about a Genotyping Experiment (MIGEN)," Standards in Genomic Sciences, 2011 A. Janssens et al "Strengthening the reporting of Genetic Risk Prediction Studies: The GRIPS statement," Genetics in Medicine,

71 " " ( I " %! %! ( ;=! Leitner F, Valencia A (2008). A text-mining perspective on the requirements for electronically annotated abstracts. Febs Letters 582(8): Lu Z, Hirschman L (2012). Biocuration workflows and text mining: overview of the BioCreative 2012 Workshop Track II. Database Vol. 2012(Article ID bas043). Biograph, BioRat 38

72 " " ( I " %! %! ( ;=! % %& /E HuGe, Ingenuity Pathway Analysis, Knome, Alamute, HuGe (W. Yu, et al: "A navigator for human genome epidemiology," Nature Genetics, 2008) The Science Behind an Answer 3:

73 " " ( I " %! %! ( ;=! %!" K. Goddard et al., "Building the evidence base for decision making in cancer genomic medicine using comparative effectiveness research," Genetics in Medicine, 2012 M. Gwinn, et al, "Horizon scanning for new genomic tests," Genetics in Medicine >

74 " " ( I " %! %! ( ;=! % %& /E!"!+- G. Shi et al: "Mining Gold Dust Under the Genome Wide Significance Level: A Two-Stage Approach to Analysis of GWAS," Genetic Epidemiology, 2011 E. Evangelou et al: "Meta-analysis methods for genome-wide association studies and beyond," Nature Reviews Genetics,

75 0 #/ ' " Bayes rule: Systems-based, Bayesian biomarker discovery p( Model Data) p( Data Model) p( Model) A priori distribution (prior), p(model): is a technical tool for inversion (to achieve a direct probabilistic statement), provides a principled solution to incorporate prior knowledge. Open problem: derivation and formalization of informative (domain/data specific) priors. Samples 7' - 05" % ). -67

76 Similarity-based fusion in drug repositioning Chemical Side-effects Target prot. MMoA Pathways Tanimoto Y Query-based optimal fusion

77 E ! 9>'" ' 555 5!! 5!"!! '

78 # 8' :' >'.' 8'! 5 " :' % >',.' I " %; %?8:* %; % 0' 2'? :: P`=P 3' 17 & & & 4' # & F% <' &, 89' " " 88' 1, 17& & 8:'