Development of multivariate classifiers in cancer


Published on

Short presentation about development of multivariate classifiers to predict chemotherapy treatment responses in breast cancer. The steps of workflow are briefly described and the results indicate that expression data on micro-RNA in breast cancer alone are not sufficient to predict treatment responses.

Published in: Health & Medicine
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Development of multivariate classifiers in cancer

  1. 1. Multivariate Algorithms and Classifiers in Cancer Micro-RNA profiles help predict distant diseasefree survival in breast cancer Bits and pieces of bioinformatics workflow Mehis Pold, MD October 18, 2013
  2. 2. Feature Selection & algorithm development Training samples Iterative process Internal Algorithm Validation Validation samples Clinical Validation Training and validation datasets in each step don’t overlap Rule of thumb: validation always produces weaker statistics than training
  3. 3. • Analysis of early primary breast cancer to identify prognostic markers and associated pathways: mRNA and miRNA profiling • GEO (Gene Expression Omnibus) accession ID: GSE22220 • Technology platform: ILLUMINA • 733 micro-RNA • 210 breast cancer samples • 79 complete pathological response (pCR) to chemotherapy; 131 recurrent disease samples (RD) • Data collected up to 10 years after start of chemotherapy Buffa et al. microRNA-Associated Progression Pathways and Potential Therapeutic Targets Identified by Integrated mRNA and microRNA Expression Profiling in Breast Cancer. Cancer Res. 2011, 71:5635
  4. 4. BIOINFORMATICS WORKFLOW Multiple statistical approaches to maximize outcome TRAINING SET: 36 RD 74 pCR VALIDATION SET: 43 RD 57 pCR Kaplan-Meier & ROC Sensitivity (Se) Specificity (Sp) Positive Predictive Value (PPV) Negative Predictive Value (NPV) Comparison of two algorithms and classification by kNN Custom-scripting (R, VBA) Standard Software : MS Excel Medical Statistics: MedCalc
  5. 5. FEATURE SELECTION Reduction of dimensionality from n = 733 to n = 1 Approach 1: iterative clustering Approach 2: T-test combined with enriching for weak inter-profile correlation Significance of feature selection evaluated by KaplanMeyer survival analysis and ROC (receiver-operator curve) RD Up pCR Down
  6. 6. KAPLAN-MEIER SURVIVAL CURVE The Kaplan–Meier estimator, also known as the product limit estimator, is an estimator for estimating the survival function from lifetime data. In medical research, it is often used to measure the fraction of patients living for a certain amount of time after treatment. In economics, it can be used to measure the length of time people remain unemployed after a job loss. In engineering, it can be used to measure the time until failure of machine parts. In ecology, it can be used to estimate how long fleshy fruits remain on plants before they are removed by frugivores. The estimator is named after Edward L. Kaplan and Paul Meier. Receiver operating characteristic (ROC) In signal detection theory, a receiver operating characteristic (ROC), or simply ROC curve, is a graphical plot which illustrates the performance of a binary classifier system as its discrimination threshold is varied. It is created by plotting the fraction of true positives out of the total actual positives (TPR = true positive rate) vs. the fraction of false positives out of the total actual negatives (FPR = false positive rate), at various threshold settings. TPR is also known as sensitivity (also called recall in some fields), and FPR is one minus the specificity or true negative rate.
  9. 9. Nearest Neighbor Classification - kNN • Based on a measure of distance between observations (e.g. Euclidean distance or one minus correlation). • k-nearest neighbor rule (Fix and Hodges (1951)) classifies an observation X as follows: – find the k closest observations in the training data, – predict the class by majority vote, i.e. choose the class that is most common among those k neighbors. Classification of data in 2D space K=3 K=5
  10. 10. SUMMARY ITERATIVE CLUSTERING TO BINARY OUTCOME TRAINING p-value Kaplan-Meier ROC AOC Sensitivity Specificity 0.0001 <.0001 0.773 72 0.67 65 0.61 0.50 0.63 65 0.51 NPV 68 0.0002 0.0024 PPV VALIDATION Kaplan-Meier ROC CLASSIFICATION kNN T-TEST ENRICHED FOR WEAK CORRELATIONS TRAINING p-value Kaplan-Meier ROC AOC Sensitivity Specificity <.0001 <.0001 0.898 83 0.624 58 0.86 0.65 0.64 56 0.35 NPV 82 0.012 0.0334 PPV VALIDATION Kaplan-Meier ROC CLASSIFICATION kNN
  11. 11. CONCLUDING REMARKS • There is no single ‘right’ approach to algorithm development. • Validation always produces weaker statistics than training. • Significance of training statistics and validation statistics are not very well correlating. • Algorithms are only as stable and significant as upstream R&D data. The better standardized and controlled the wetbench, the more stable and significant the algorithms and eventual clinical validation.