Download presentation
Presentation is loading. Please wait.
1
Prediction of Optimal Cancer Drug therapies via SVM
Presented by: Xuwen Zhao
2
Overview Why we need this prediction Algorithms used
SVM (support vector machine) RFE (recursive feature elimination) 3 different conditions to test for accuracies
3
Why we need this prediction?
Part of Precision medicine Ideal form of this prediction E.g. CML (chronic myelogenous leukemia) Effectiveness in saving money and time Picture from:
4
Algorithms Applied Support Vector Machine (SVM) model Predictive model
Built with gene expression and drug sensitivity profiles Database: NCI-60 panel of human cancer cell lines Predictive model Built for seven often used drugs in ovarian cancer treatment Recursive Feature Elimination (RFE) selection Optimally distinguish cells’ sensitivity towards drugs Picture from:
5
Algorithms Applied SVM-RFE
Works in recursive manner until minimal subset of features is identified Minimal subset of features that essential to maintain optimal accuracy The lowest ranked 100 probes will be removed prior to the next round of selection in the RFE process. The process proceeds recursively until 100 features (remain). Thereafter, one feature is removed at each round of selection until a set of optimally predictive features (probes) is established. The process ends when a minimum of 10 optimally predictive features (probes) is identified.
6
A minimal of 10 optimally features obtain
Highest Rank Highest Rank 100 FEATURES REMIAN Repeat until Lowest 100 removed Lowest Rank Probe Removed Lowest Rank A minimal of 10 optimally features obtain
7
Test for Accuracies under 3 Different Conditions
Condition1: Across Variety of cancer types Result: Improves predictive accuracy Condition2: Averaging microarray probe set expression Result: Reduces predictive accuracy Condtion3: Pre-filtering learning datasets with current biological understanding
8
Condition1: Across Variety of cancer types
Why? Development of cancer progression is not only identified by tumor’s origin Gene expression with particular cancer type may underline other cancer development in other types Comparison between 2 SVM models test on Carboplatin 18 Cells lines from NCI60
9
Condition1: Across Variety of cancer types
Randomly selected to be representative of all 9 cancer types Only lung and melanoma Test on Carboplatin 75% Test on Carboplatin 87.5% Fig1. Comparison of predictive accuracy (ROC curve) for two SVM models
10
Condition2: Averaging microarray probe set expression
Example: Affymetrix Incorporate multiple probe sets per gene, thereby providing the possibility of monitoring differences in levels of alternative splicing and other post- transcriptional expression variants Use breast cancer cell lines Test on drug doxorubicin Average Affymetrix V.S. original gene expression datasets Picture from:
11
Condition2: Averaging microarray probe set expression
Averaging expression value a Original expression value c Figure 2 (a, b, c). Comparison of 2 SVM models predictive accuracies towards drug doxorubicin
12
Condtion3: Pre-filtering learning datasets with current biological understanding
Data used: breast cancer cell lines Filtering data with previously identified genes Decreases diversity in data pool 297 genes previously implicated in cancer V.S. all significant expressed genes Figure 3. ROC curves of two SVM models showing reduces predictive accuracy in pre-filtering dataset
13
Model Application on OC patients
Gene expression of 273 OC patients tumor from GEO repository Predict response to 7 often used drugs Example: Predictions of 2 individual OC patients a b Figure 4 (a, b). Predictive response of two individual OC patients.
14
Model Application on OC patients
Drug Predicted response rate Observed literature response rate Carboplatin 75% Paclitaxel 56% Docetaxel 36% 28% Cisplatin 58% 60% Gefitinib 2% 6% Doxorubicin 5% 20% Gemcitabine 47% Table 1. Comparison between predicted response rates of 273 ovarian cancer patients to 7 chemotherapeutic drugs and response rates as reported in the literature.
15
Data selection SVM-RFE Limitation Diversity Observed data
Some Comments Data selection SVM-RFE Limitation Diversity Observed data
16
Reference Cai Huang, Roman Mezencev, John F. McDonald, Fredrik Vannberg. Open source machine-learning algorithms for the prediction of optimal cancer drug therapies. PLoS One Oct.26
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.