How long can cancer patients live? The prospect of survival for cancer patients depends on some key factors, such as the age at diagnosis, cancer type, cancer stage, cancer metastasis, and genetic features. As a result, the signature identification of biomarkers, such as genes and non-coding RNAs (ncRNAs), associated with survival time is urgently required for precision medicine. A biomarker signature is defined as a minimal subset of biomarkers that are maximally predictive with respect to survival time when they work cooperatively. Biostatistics approaches are commonly used for discovery of individual biomarkers, while machine learning approaches are appropriate for identifying biomarker signatures and establishing prediction models from a large dataset, e.g., expression profiles of genes/ncRNAs. However, the training expression profiles of cancer patients are usually characterized by high-dimensional samples and small datasets, which result in both curse of dimensionality and underdetermined problems.

We propose a novel evolutionary learning platform on complex biomedical data for identification of biomarker signatures while considering the underdetermined problem. For example, the proposed method identified 18 out of 332 miRNAs and achieved a correlation coefficient of 0.88 ± 0.01 and mean absolute error of 0.56 ± 0.03 years between real and estimated survival time of lung adenocarcinoma patients. The platform has been validated using miRNA/lncRNA expression profiles of patients with lung adenocarcinoma, glioblastoma multiforme, breast cancer, ovarian cancer and neuroblastoma from The Cancer Genome Atlas and Gene Expression Omnibus databases. The analysis of all signatures of genes/ncRNAs for cancer patients can help to develop drugs and gene targeted therapy.

Figure 1: Evolutionary learning platform for biomedical data.

Figure 2: The computer aided diagnosis system won an Entrepreneurial Potential Award in the FITI Program Competition of the Ministry of Science and Technology.