Introduction to Data Mining / Bioinformatics
3 likes2,323 views
This document discusses the principles and techniques of data mining, particularly in the field of bioinformatics. It covers the procedure of data mining, which includes acquisition, preprocessing, feature selection, classification, validation, and prediction. The lecture aims to help understand relationships between measurable properties and important biological outcomes, emphasizing model validation and iterative improvements based on predictions.
Introduction to Data Mining / Bioinformatics
- 1. Introduction to Bioinformatics: Mining Your Data Gerry Lushington Lushington in Silico modeling / informatics consultant
- 2. What is Data Mining? Use of computational methods to perceive trends in data that can be used to explain or predict important outcomes or properties Applicable across many disciplines: Molecular bioinformatics Medical Informatics Health Informatics Biodiversity informatics
- 3. Example Applications: Find relationships between: Convenient Observables vs. Important Outcomes a) Relative gene expression data 1. Disease susceptibility b) Relative protein abundance data 2. Drug efficacy c) Relative lipid & metabolite profiles 3. Toxin susceptibility d) Glycosylation variants 4. Immunity e) SNPs, alleles 5. Genetic disorders f) Cellular traits 6. Microbial virulence g) Organism traits 7. Species adaptive success h) Behavioral traits 8. Species complementarity i) Case history
- 4. Goals for this lecture: Focus on Data Mining: how to approach your data and use it to understand biology Overview of available techniques Understanding model validation Try to think about data you’ve seen: what techniques might be useful? Don’t worry about grasping everything: K-INBRE Bioinformatics Core is here to help!!
- 5. Basic Data Mining: Find relationships between: a) Easy to measure properties vs. b) Important (but harder to measure) outcomes or attributes Use relationships to understand the conceptual basis for outcomes in b) Use relationships to predict outcomes in new cases where outcome has not yet been measured
- 6. Basic Data Mining: simple measureables
- 7. Basic Data Mining: general observation Unhappy Happy
- 8. Basic Data Mining: relationship (#1) Unhappy Happy Blue = happy; Red = unhappy accuracy = 12/20 = 60%
- 9. Basic Data Mining: relationship (#2) Unhappy Happy Blue + BIG Red = happy; little red = unhappy accuracy = 17/20 = 85%
- 10. Data Mining: procedure 1. Data Acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration
- 11. Data Mining: procedure 1. Data acquisition 2. Data Preprocessing Peak heights? 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration Peak positions? Key issues include: a) format conversion from instrument b) any necessary mathematical manipulations (e.g., Density = M/V)
- 12. Data Mining: procedure 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration Key issues include: a) Normalization to account for experimental bias b) Statistical detection of flagrant outliers
- 13. Data Mining: procedure 1. Data acquisition 2. Data Preprocessing 3. Feature Selection Use controls to 4. Classification scale data 5. Validation 6. Prediction & Iteration Key issues include: a) Normalization to account for experimental bias b) Statistical detection of flagrant outliers C C 1 2 3 C 1 2 3 C 1 2 3 C 1 2 3
- 14. Data Mining: procedure 1. Data acquisition 2. Data Preprocessing 3. Feature Selection Subjective 4. Classification (requires experience 5. Validation and/or domain 6. Prediction & Iteration knowledge) Key issues include: a) Normalization to account for experimental bias b) Statistical detection of flagrant outliers
- 15. Data Mining: procedure 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration Which out of many measurable properties relate to outcome of interest? a) Intrinsic information content b) Redundancy relative to other properties c) Correlation with target attribute d) Iterative model training
- 16. Data Mining: procedure 1. Data acquisition 2. 3. Data Preprocessing Feature Selection x x 4. Classification 5. Validation 6. Prediction & Iteration 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 Which out of many measurable properties relate to outcome of interest? a) Intrinsic information content b) Redundancy relative to other properties c) Correlation with target attribute d) Iterative model training
- 17. Data Mining: procedure 1. Data acquisition 2. 3. Data Preprocessing Feature Selection x 4. Classification 5. Validation 6. Prediction & Iteration 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 Which out of many measurable properties relate to outcome of interest? a) Intrinsic information content b) Redundancy relative to other properties c) Correlation with target attribute d) Iterative model training
- 18. Data Mining: procedure 1. Data acquisition 2. 3. Data Preprocessing Feature Selection x 4. Classification 5. Validation 6. Prediction & Iteration 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 Which out of many measurable properties relate to outcome of interest? a) Intrinsic information content b) Redundancy relative to other properties c) Correlation with target attribute d) Iterative model training 1 2 3 4
- 19. Data Mining: procedure 1. Data acquisition • Train preliminary models based on random sets of properties 2. Data Preprocessing • Evaluate models according to 3. Feature Selection correlative or predictive performance 4. Classification • Experiment with promising sets adding 5. Validation or deleting descriptors to gauge impact 6. Prediction & Iteration on performance Which out of many measurable properties relate to outcome of interest? a) Intrinsic information content b) Redundancy relative to other properties c) Correlation with target attribute d) Iterative model training
- 20. Data Mining: procedure 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration Predict which sample will have which outcome? a) Correlative methods b) Distance-based clustering c) Boundary detection d) Rule learning e) Weighted probability
- 21. Data Mining: procedure y 1. Data acquisition 2. Data Preprocessing x 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration Predict which sample will have which outcome? a) Correlative methods b) Distance-based clustering c) Boundary detection d) Rule learning e) Weighted probability
- 22. Data Mining: procedure y 1. Data acquisition 2. Data Preprocessing x 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration -n y +n Predict which sample will have which outcome? NO YES a) Correlative methods b) Distance-based clustering c) Boundary detection d) Rule learning e) Weighted probability
- 23. Data Mining: procedure x2 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration x1 Predict which sample will have which outcome? a) Correlative methods b) Distance-based clustering c) Boundary detection d) Rule learning e) Weighted probability
- 24. Data Mining: procedure y1 y2 x2 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification y3 5. Validation y4 6. Prediction & Iteration x1 Predict which sample will have which outcome? a) Correlative methods y1 = resistant to types I & II diabetes b) Distance-based clustering y2 = susceptible only to type II c) Boundary detection d) Rule learning y3 = susceptible only to type I e) Weighted probability y4 = susceptible to types I & II
- 25. Data Mining: procedure Resistant to type I x2 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration x1 Susceptible to type I Predict which sample will have which outcome? a) Correlative methods b) Distance-based clustering c) Boundary detection d) Rule learning e) Weighted probability
- 26. Data Mining: procedure Resistant to type I x2 1. Data acquisition 2. Data Preprocessing b 3. Feature Selection 4. Classification a 5. Validation 6. Prediction & Iteration c x1 Susceptible to type I Predict which sample will have which outcome? a) Correlative methods b) Distance-based clustering If x1 < c and x2 > a then resistant c) Boundary detection Else if x1 > c and x2 > b then resistant d) Rule learning Else susceptible e) Weighted probability E=9
- 27. Data Mining: procedure Resistant Susc. 1. Data acquisition 2. Data Preprocessing 3. Feature Selection a x1 4. Classification 5. Validation Susc. Resistant 6. Prediction & Iteration b x2 Predict which sample will have which outcome? a) Correlative methods Resistant Susc. b) Distance-based clustering c) Boundary detection c Fx1 - d) Rule learning Gx2 e) Weighted probability If Fx1 - Gx2 < c then resistant Else susceptible
- 28. Data Mining: procedure 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration Define criteria and tests to prove model validity a) Accuracy b) Sensitivity vs. Specificity c) Receiver Operating Characteristic (ROC) plot d) Cross-validation
- 29. Data Mining: procedure x2 1. Data acquisition 2. Data Preprocessing 3. Feature Selection Resistant (Neg.) 4. Classification 5. Validation Susc. 6. Prediction & Iteration x1 (Pos.) Define criteria and tests to prove model validity a) Accuracy Accuracy = (TP + TN) b) Sensitivity vs. Specificity TP + TN + FP + FN c) Receiver Operating Characteristic (ROC) plot d) Cross-validation = 142 / 154
- 30. Data Mining: procedure x2 1. Data acquisition 2. Data Preprocessing 3. Feature Selection Resistant (Neg.) 4. Classification 5. Validation Susc. 6. Prediction & Iteration x1 (Pos.) Define criteria and tests to prove model validity a) Accuracy Sensitivity = TP = 67 / 72 b) Sensitivity vs. Specificity c) Receiver Operating Characteristic (ROC) plot TP + FN d) Cross-validation FPR = FP = 6 / 81 TN + FP Note: Specificity = 1 - FPR
- 31. Data Mining: procedure x2 1. Data acquisition 2. Data Preprocessing 3. Feature Selection Resistant (Neg.) 4. Classification less 5. Validation Varying Susc. 6. Prediction & Iteration model more x1 (Pos.) stringency Define criteria and tests to prove model validity a) Accuracy Sensitivity = TP = 69 / 72 b) Sensitivity vs. Specificity c) Receiver Operating Characteristic (ROC) plot TP + FN d) Cross-validation FPR = FP = 19 / 81 TN + FP Note: Specificity = 1 - FPR
- 32. Data Mining: procedure Sens 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration FPR Define criteria and tests to prove model validity a) Accuracy b) Sensitivity vs. Specificity c) Receiver Operating Characteristic (ROC) plot d) Cross-validation
- 33. Data Mining: procedure Sens 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration FPR Define criteria and tests to prove model validity Area under curve is a) Accuracy excellent measure of b) Sensitivity vs. Specificity model performance c) Receiver Operating Characteristic (ROC) plot d) Cross-validation 1.0: perfect model 0.5: random
- 34. Data Mining: procedure 1. Data acquisition Predictions are imperfect due to: 2. Data Preprocessing • Imperfect Algorithms 3. Feature Selection • Imperfect Data 4. Classification 5. Validation 6. Prediction & Iteration Define criteria and tests to prove model validity a) Accuracy b) Sensitivity vs. Specificity c) Receiver Operating Characteristic (ROC) plot d) Cross-validation
- 35. Cross-Validation: • Carefully monitor features that are useful across different independent data subsets • This can be accomplished with N-fold cross-validation: Trial 1 Trial 2 Trial 3 Trial 4 Trial 5 Test Train Model performance = mean predictive performance over 5 trials • Best feature selection and classification algorithms will yield best consistent performance across independent trials • Best features will be consistently important across trials
- 36. Data Mining: procedure 1. Data acquisition 2. Data Preprocessing 3. Feature Selection 4. Classification 5. Validation 6. Prediction & Iteration Analysis is only useful if it is used; only improves if it is tested a) Good validation requires successful new predictions b) Imperfect predictions can lead to method refinement and greater understanding
- 37. Questions? Lushington in Silico Geraldlushington3117 at aol.com Geraldlushington.org