Rule learning for gene expression data

Transcription

1 Rule learning for gene expression data Stefan Enroth Original slides by Torgeir R. Hvidsten The Linnaeus Centre for Bioinformatics

2 Predicting biological process from gene expression time profiles Papers: I. T. R. Hvidsten, A. Lægreid and J. Komorowski. Learning rule-based models of biological process from gene expression time profiles using gene ontology, Bioinformatics 19(9): , II. A. Lægreid, T. R. Hvidsten, H. Midelfart, J. Komorowski and A. K. Sandvik. Predicting Gene Ontology Biological Process From Temporal Gene Expression Patterns, Genome Research, 13(5): , 2003.

3 Hierarchical clustering 3 Iyer et al., The transcriptional program in the response of human fibroblasts to serum, Science, 283(5398): 83-87, 1999

4 4 Gene Ontology vs. expression clustering

5 Methodology Ontology Process 1. Annotation Transport Defense response Positive control of cell proliferation Cell cycle control g 2... g 2... g 4... g 5 g Extracting features for learning Gene 0HR 15MIN 30MIN 1HR 2HR 4HR 6HR 8HR 12HR 16HR 20HR 24HR Process g Unknown g Transport and defense response g Cell cycle control Positive control of g cell proliferation g Positive control of cell proliferation Inducing minimal decision rules using rough sets (Increasing) AND 6-10(Decreasing) AND 14-18(Constant) => GO(cell proliferation) 5 4. The function of uncharacterized! genes is predicted using the rules

6 Rule Induction IF THEN 0-4 (Constant) AND 0-10(Increasing) GO(prot. met. and mod.) OR GO(mesoderm develop.) OR GO(prot. biosynt.) 6 IF-part (antecedent, premise) the minimal set of discrete changes in expression needed to uphold the discriminatory power of the full data set THEN-part (conclusion) all functions of genes described by the antecedent-side We want rules that describe the expression profiles of several genes with one or a few functions accuracy: the fraction of genes matching the IF-part that are annotated with the process in the THEN-part coverage: the fraction of genes annotated with the process in the THEN-part that matches the IF-part

7 Rule example Rule 0-4(Constant) AND 0-10(Increasing) => GO(protein metabolism and modification) OR GO(mesoderm development) OR GO(protein biosynthesis) Covered genes 4: M35296 J02783 D13748 X : X : D

8 Classification X Function? 8 IF THEN IF THEN IF THEN IF THEN IF THEN IF THEN IF THEN IF THEN IF THEN IF THEN Votes are normalized and processes with vote fractions higher than a selection-threshold are chosen as predictions IF 0-4(Constant) AND 0-10(Increasing) THEN GO(protein metabolism and modification ) OR GO(mesoderm development) OR GO(protein biosynthesis) +1 Process Votes protein metabolism and modification 6 mesoderm development 3 proteolysis and peptidolysis 2 transcription 1 protein biosynthesis 1 vision 1

9 Cross validation Iteration 1 Iteration 2 Iteration 3 Observation 1 Observation 2 9 Fold 1 Fold 2 Fold 3 Training set Training set Test set Training set Test set Training set Test set Training set Training set Observation n

10 Evaluation Evaluation technique divide examples into training set and test set cross validation Evaluation measures: accuracy = (TP+TN)/(TP+FN+TN+FP) sensitivity = TP/(TP+FN) specificity = TN/(TN+FP) Confusion matrix: 10

11 Threshold selection Gene with function protein biosynthesis Gene with a different function sensitivity: TP/(TP+FN) specificity: TN/(TN+FP) 1 Fraction of votes for protein biosynthesis Threshold 1 Threshold 2 g 1 g 2 g 3 g 4 g 5 g 6 g 7 g 8 g 9 g 10 g 11 g 12 g 13 Test set Sensitivity = 2/3, Specificity=1 Sensitivity = 1, Specificity=2/3

12 ROC analysis and classifier evaluation 1 sensitivity Perfect discrimination AUC No discrimination ROC: Receiver Operating Characteristics curve results from plotting sensitivity against specificity for all possible thresholds sensitivity: TP/(TP+FN) specificity: TN/(TN+FP) AUC: Area Under the ROC Curve specificity False alarm 1

13 ROC analysis and classifier evaluation 1 Perfect discrimination A B Which ROC curve is better? A dominants B and C and clearly has a higher AUC sensitivity C No discrimination B and C have approximately the same AUC B is better for some thresholds, C for others specificity 1

14 Cross validation estimates PROCESS AUC SE P-VALUE Ion homeostasis Protein targeting Blood coagulation DNA metabolism Intracellular signaling cascade Cell cycle Energy pathways Oncogenesis Circulation Cell death Developmental processes Defense (immune) response Transcription Cell adhesion Stress response Protein metabolism and modification Cell motility Cell surface rec linked signal transd Lipid metabolism Cell organization and biogenesis Cell proliferation Transport Amino acid and derivative metabolism AVERAGE Over all classes: Coverage (recall) = TP/(TP+FN) Precision = TP/(TP+FP) Coverage: 84% Precision: 50% Coverage: 71% Precision: 60% Coverage: 39% Precision: 90% *Iyer et al. 14

15 Discovering regulatory binding site modules Paper: III. T. R. Hvidsten, B. Wilczyński, A. Kryshtafovych, J. Tiuryn, J. Komorowski and K. Fidelis. Discovering regulatory binding site modules using rule-based learning, Genome Research 15: , 2005

16 Gene regulation Gene expression is regulated by regulatory proteins (transcription factors) Transcription factors depend on recognizing sequence motifs (binding sites) in order to effect the expression of genes Transcription factors combine to respond to a large number of stress factors (e.g. heat shock) with a large number of expression outcomes Enhancer Silencer Response elements Promoter Regulatory region Binding sites Transcription region Yeast 16 Promoter

17 Background Many studies have used gene expression data to search for overrepresented sequence motifs in co-expressed genes Pilpel et al. (2001) found that genes sharing pairs of binding sites are significantly more likely to be co-expressed than genes with only single binding sites in common Expression coherence score (EC) 17 Pilpel, Y., P. Sudarsanam, and G.M. Church Identifying regulatory networks by combinatorial analysis of promoter elements. Nat Genet 29:

18 Introductory remarks We want to explore the combinatorial nature of gene regulation ASSUMPTION: co-regulated genes (genes regulated by the same transcription factors through the same binding sites) are co-expressed (have similar expression profiles) Genome-wide analysis of combinatorial regulation using sequence and expression data 18 Binding site module: a set of binding sites that is used to co-regulate a set of genes

19 Data and Method Database of 43 known and 313 putative yeast binding site motifs Expression profiles for yeast genes measured under six different conditions: cell cycle and five stress conditions sporulation, diauxic shift, heat shock and cold shock, pheromone and DNA-damaging agents Gene HAP234 Binding sites RAP1 PHO SWI5 ECB MCM1' RPL18A RPS18A RPL16B RPL26A RPS24A RPL RPL14A SST DRS GIT CLN RPO BIT Next gene Similar expression to RPL18A? Rule learning* yes yes yes yes yes yes yes no no no no no no IF RAP1 AND SWI5 AND MCM1' THEN 4 Evaluation: Gene Ontology Binding data Filtering*

20 An example of a binding site module IF RAP1 AND SWI5 AND MCM1' THEN The rule was (re-) discovered in five of the six expression data sets a) Cell cycle: RPL30, RPL18A, RPL26A, RPL14A, RPS18A, (SST2), RPL16B b) Sporulation: RPL30, RPL18A, RPL14A, (RPS18A), RPL16B, RPL26A The central gene in the expression cluster is underlined c) Diauxic shift: RPL30, RPL18A, RPL14A, RPS18A, SST2, RPL16B, RPL26A Genes with differing expression profiles are in parentheses d) Heat and cold shock: RPS24A, RPL26A, RPL14A, RPS18A, RPL16B e) DNA-damaging agents: RPL30, RPL18A, RPL26A, RPL14A, RPS18A, (SST2), RPL16B

21 Evaluation IF RAP1 AND SWI5 AND MCM1' THEN <similar expression> Gene symbol Biological process Molecular function Cellular component Possible transcription factors (P-value < 0.01) RPL16B protein biosynthesis RNA binding, structural constituent of ribosome cytosolic ribosome (sensu Eukarya), large ribosomal subunit FHL1, GAT3, PDR1, RAP1, RGM1, YAP5 RPL26A protein biosynthesis RNA binding, structural constituent of ribosome cytosolic ribosome (sensu Eukarya), large ribosomal subunit FHL1,RAP1 RPS18A protein biosynthesis structural constituent of ribosome cytosolic ribosome (sensu Eukarya), eukaryotic 43S preinitiation complex,eukaryotic 48S initiation complex, mall ribosomal subunit FHL1, GAT3, HIR2, RAP1, RGM1, YAP5 RPL30 protein biosynthesis, rrna processing, mrna splicing, regulation of translation structural constituent of ribosome cytosolic ribosome (sensu Eukarya), cytoplasm, large ribosomal subunit FHL1, GAT3, RAP1, SFP1 RPL18A protein biosynthesis structural constituent of ribosome cytosolic ribosome (sensu Eukarya), large ribosomal subunit FHL1, MAL13, RAP1, YAP5 RPL14A protein biosynthesis RNA binding, structural constituent of ribosome cytosolic ribosome (sensu Eukarya), large ribosomal subunit FHL1, GAT3, GRF10(Pho2), GTS1, RAP1 SST2 signal transduction, adaptation to pheromone during conjugation with cellular fusion GTPase activator activity plasma membrane DIG1, FHL1, RAP1, STE12 RPS24A protein biosynthesis structural constituent of ribosome cytosolic ribosome (sensu Eukarya), eukaryotic 43S pre-initiation complex, eukaryotic 48S initiation complex, small ribosomal subunit FHL1, GAT3, PDR1, RAP1, RGM1, SMP1, YAP5 21 P-VALUE 2.35E-04 (protein biosynthesis) 2.36E-06 (structural constituent of ribosome) 5.66E-07 (cytosolic ribosome) 2.38E-10 (FHL1), 3.38E-08 (RAP1), 2.96E-05 (GAT3)

22 Model-based detection of periodic expression Paper: IV. C.R. Andersson, T.R. Hvidsten, A. Isaksson, M.G. Gustafsson and J. Komorowski. Revealing cell cycle control by combining model-based detection of periodic expression with novel cisregulatory descriptors, BMC Systems Biology, 1: 45, 2007.

23 Experiment Induction Prior knowledge Hypothesis Experiment Induction A technique that infers generalizations from the information in the data. Inference Inference the reasoning involved in drawing a conclusion or making a logical judgment on the basis of circumstantial evidence and prior conclusions rather than on the basis of direct observation 23

24 Synchronization To recover mrna-levels, cultures must be synchronized. Synchronization halts cells at a particular point in the cell cycle. Typically the mating system or temperature sensitive mutants are used (cdc28, cdc15). Spellman (1998). S cerevisae under various synchronizations (alpha-factor, ts cdc15, cdc28 and elutriation) Synchronization reveals periodic expression Periodic expression is related to the cell cycle 24

25 Detecting periodically expressed genes Signal Expression Time Which model is most similar to the signal? => Probability (Periodical expression) Model 0 Model 1 Expression Expression 25 Time Time T

26 The models H 0 : y t = μ + ε H = μ + ω + φ 1 : y t A cos( t ) + ε Prior knowledge: The period T 26

27 27 Conditional periodicity

28 28 Method

29 Comparison to clustering Known cell cycle related: sequence motifs CCA, ECB, MCB, SWI5, SCB, SFF, MCM1, SFF' and MCM1. transcription factors ACE2, FKH1, FKH2, GTS1, HIR1, HIR2, MBP1, MCM1, NDD1, STB1, SWI4,SWI5, SWI6, XBP1, YBR267W, YHP1 and YOX1. 29

30 Examples of interactions The point (0.034, 0.73) on the curve is associated with the 145 rules with p-value lower than These rules include 19 of the 26 known phase specific regulators (73%) and 18 other regulators (3.4%). Furthermore, they describe 24% of the genes in the periodic classes. 30

31 Examples of interactions Ellipses/rectangles: transcription factors/sequence motifs Green/Blue: Cell cycle related transcription factors/sequence motifs Red: interactions between transcription factors 31

32 Summary Gene expression data can be used in many types of applications in which rule-based methods can be used to synthesis hypothesis. Discovered rules/classes must always be validated against real world knowledge! If appropriate, prior real world knowledge can be incorporated into the models. 32