A method to improve survival prediction using mutual information based network
Download as PPTX, PDF1 like313 views
The document presents a method to enhance survivability prediction in cancer patients using a network-constrained cox regression (NetCox) model that incorporates mutual information derived from gene co-expression networks. It assesses the predictive power of various genomic profiles (CNA, mRNA, methylation) with a focus on mutual information and performance metrics like time-dependent ROC and log-rank tests on datasets involving 10,022 genes from 340 patients. Results indicate the importance of network information in determining patient risk classifications and highlight specific gene expression characteristics linked to clinical outcomes.
A method to improve survival prediction using mutual information based network
- 1. A method to improve predictivity of Net-Cox using Mutual Information based network
- 2. Abstract Cancer genes Gene co-expression Network NetCox (Network-constrained Cox regression) Predict Survivability 1g 2g 3g Dataset Correlation between gene pairs
- 3. Abstract CNA mRNA METH TCGA data 10,022 genes 340 patients Censored : 0 Observed : 1 Clinical outcome M.I. based Network NetCox (Network-constrained Cox regression) Predict Survivability 3g 1g 2g 𝐼(𝑔𝑖, 𝑔𝑗 ; 𝑌) Mutual Information between gene pairs and clinical outcome of patients
- 4. M.I. based network gene pair clinical outcome of patients Joint entropy of X, YEntropy of X Mutual information of X, Y Mutual information of two genes and clinical outcome of patient 0, 0.1, 0.2, …, 0.7 Threshold from permutation strategy 1g 2g 3g
- 5. NetCox (Network-constrained Cox regression) Network information 1e-4, 1e-3, 1e-2, 1e-1 0.1, 0.3, 0.5, 0.7, 0.9, 1 L2-Cox The importance of network information increases Regression coefficients Prognostic index Profile data Which parameter is the best? Predict Survivability
- 6. Cross validation for optimal parameters profile G 10022 10022 network profileX Data 340 patients 10022 genes 1 1 0 1 … 340 patients Clinical outcome uncensored (deceased) censored (living) Test set Training set 5-fold cross validation Sort in ascending order by survival time of patients test profileX NetCox training 'XPI
- 7. Evaluation 40% 40% High-risk group low-risk group PIrank the patients • Log-rank test • Time dependent ROC 𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 𝑐, 𝑡 𝑓 𝑋 = Pr 𝑓 𝑋 > 𝑐|𝛿 𝑡 = 1 , 𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 𝑐, 𝑡 𝑓 𝑋 = Pr 𝑓 𝑋 ≤ 𝑐|𝛿 𝑡 = 0 • 𝑓 𝑋 = 𝑋′ 𝛽 • 𝑐 ∶ 𝑐𝑢𝑡𝑜𝑓𝑓 𝑝𝑜𝑖𝑛𝑡 • 𝑡 ∶ 𝑡𝑖𝑚𝑒 • 𝛿 𝑡 : 𝑒𝑣𝑒𝑛𝑡 𝑖𝑛𝑑𝑖𝑐𝑎𝑡𝑜𝑟 𝑎𝑡 𝑡𝑖𝑚𝑒 𝑡 The area under the ROC curve for any time t -> how well PI classifies the patients into high or low risk group −log(𝑝𝑣𝑎𝑙𝑢𝑒) Evaluate whether the patients are assigned to the right group.
- 8. Results - Time-dependent ROC & log-rank test
- 9. Results Mutual Information Correlation Functional Linkage L2-Cox CNA 0.6276 0.6265 0.6192 0.6229 mRNA 0.6427 0.6481 0.6366 0.6418 METH 0.5876 0.5845 0.5855 0.5663 Mutual Information Correlation Functional Linkage L2-Cox CNA 1.4068 1.4335 1.0968 1.2020 mRNA 1.8107 2.0748 1.2346 1.5018 METH 0.9714 0.8576 0.7596 0.7407 (a) Best Performance Comparison with time AUC (b) Best Performance Comparison with log-rank test
- 10. Results
- 11. Results Genomic Profile Test lambda alpha theta threshold No. of gene pairs above threshold Percentage CNA log-rank 1E-03 0.3 0.4 0.0930 66,754 0.13% AUC 0.7 0.3 0.0863 179,963 0.36% mRNA log-rank 1E-04 0.1 0.6 0.1221 6,487 0.01% AUC METH log-rank 1E-04 0.1 0.3 0.1017 32,282 0.06% AUC 0.3 0.5 0.1173 3,641 0.01%
- 12. Results – CNA network
- 13. Results – mRNA network
- 14. Results – Methylation network
- 15. Results
- 16. Results Profile Category ID description count p-value corr p-value CNA BP 42335 cuticle development 2 1.91E-06 2.53E-03 CNA MF 16174 NAD(P)H oxidase activity 2 2.26E-05 5.13E-03 meth BP 97479 synaptic vesicle localization 4 1.86E-05 5.82E-03 meth BP 48489 synaptic vesicle transport 4 1.74E-05 5.82E-03 mRNA BP 6915 apoptotic process 16 1.55E-07 1.81E-04 mRNA BP 30162 regulation of proteolysis 10 1.04E-06 3.87E-04 P-value < 0.01