Modern bioinformatics offers more and more offending challenges that came from highly through-output biomedical and genome data (microarrays) received by immunoscreening of cancer patients and healthy donors. Highly dimensional and much less impressing in sample size data sets experiencing unpredictable and unstable distributions for their covariates make it impossible to use lots of classifiers and regression tools that confirmed their efficiency for more simple and controllable tasks. Improvement of kernel methods for Support Vector Machine (SVM) classification in cancer diagnostics is the main intent of this paper. Unfortunately microarray data causes significant decay in effect of structured risk minimization on examined data. This paper proposes new more effective method for learning SVM kernel on noisy and unpredictable by their nature data sources. As result of this paper we have inferred new method for learning optimal kernel hyperparameters using Cox regression.