Fig 1.
An SVM-RFE predictive model of carboplatin sensitivity for NCI-60 cell lines.
(A) Ranked display of -log transformed GI50 values for carboplatin for each of the NCI-60 cell lines. Blue circles = carboplatin resistant cells; red circles = carboplatin sensitive cell lines. Cell lines with GI50 values within ±0.5 SD of the mean (green circles) are less reliably classified as resistant or sensitive and were, thus, not employed in learning datasets. Test sets were selected from cell lines across the entire distribution; (B) Evolution of accuracy of predicted response to carboplatin using SVM-RFE selection for gene probe classifiers; (C) Visualization of the optimal separation between carboplatin sensitive and resistant NCI-60 cell lines. The X-axis is the optimal weight vector (prediction score) of the SVM model for carboplatin; the Y-axis is the -log transformed GI50 values for carboplatin.
Fig 2.
The influence of learning datasets on the predictive accuracy of SVM-RFE models.
(A) Comparison of predictive accuracy (ROC curves) for two SVM models of response to carboplatin using a learning dataset derived from 2 cancer types (lung, melanoma) vs. 9 cancer types (brain, breast, lung, leukemia, renal, colon, ovarian, prostate and melanoma). In each case, the data were derived from a total of 18 cell lines. The results indicate that the model built using learning set data from 9 cancer types generates a more accurate prediction (see also Fig D in S1 File); (B,C,D) Prediction of the sensitivity of breast cancer cell lines to doxorubicin. In one case, the model was built using a learning dataset comprised of average gene expression values. In the other case, the model was built using a learning dataset comprised of the expression values of all gene probes. The results demonstrate that the model built using probe set data is more accurate than the model built using average gene expression data; (C) prediction score accuracy using average gene expression values; (D) prediction score accuracy using expression values of all gene probes (Red circles = drug sensitive training set; Blue circles = drug resistant training set; Black diamonds = breast cancer cells test set).
Fig 3.
Pre-filtering of learning datasets can reduce the accuracy of predictive models.
Shown is the predicted sensitivity of breast cancer cell lines to doxorubicin by two SVM models built using different learning datasets. In one case, the model was built using a learning dataset limited to the expression of 297 genes previously associated with cancer onset/progression [19]. In the other case, the model was built using a learning dataset drawn from all significantly expressed genes (Table A in S2 File). The results indicate that pre-filtering of the learning dataset to only include gene expression values of previously identified cancer related genes reduces predictive accuracy. (A) Quadrant plot of SVM predicted sensitivity to doxorubicin vs. observed sensitivity to doxorubicin of model built using a learning dataset pre-filtered for genes previously associated with cancer onset/progression; (B) Quadrant plot of SVM predicted sensitivity to doxorubicin vs. observed sensitivity to doxorubicin of model built using all gene expression data (Table A in S2 File); (C) ROC curves of the two models showing reduced predictive accuracy associated with the pre-filtered learning dataset (Red circles = drug sensitive training set; Blue circles = drug resistant training set; Black diamonds = breast cancer cells test set).
Fig 4.
Individual and aggregate prediction of response to chemotherapeutic drugs.
The SVM algorithms output binary classifications for each drug (sensitive/resistant) established through a decision function that numerically separates cancer cells predicted to respond to the drug (positive score) from those predicted to be non-responders (negative score). (A) The predicted response of an individual patient (GSM516724) to seven chemotherapeutic drugs. This patient is predicted to respond favorably to the first line therapies of carboplatin (score 2.88) and paclitaxel (score 3.20). (B)The predicted response of a second individual OC patient (GSM516801) to seven chemotherapeutic drugs. The patient is predicted NOT to respond favorably to the first line therapies of carboplatin (score -0.28) and paclitaxel (score -2.53). (C) Density plot of aggregate prediction scores for 3 GEO data sets of 273 ovarian cancer patients and the predicted group response rate for each drug. (D) Scatter plot of the predicted group response rates vs. the observed group responses of OC patients to seven chemotherapeutic drugs (Linear regression p value = 0.0031, R2 = 0.8201) (Table F in S2 File).
Fig 5.
Pseudo code for the RFE approach.
This approach takes the microarray expression data of NCI-60 cancer cell lines as input data, and the output is a model with the most informative features.