Use of Artificial Intelligence and Machine Learning Algorithms with Gene Expression Profiling to Predict Recurrent Nonmuscle Invasive Urothelial Carcinoma of the Bladder.

PURPOSE: Due to the high recurrence risk of nonmuscle invasive urothelial carcinoma it is crucial to distinguish patients at high risk from those with indolent disease. In this study we used a machine learning algorithm to identify the genes in patients with nonmuscle invasive urothelial carcinoma at initial presentation that were most predictive of recurrence. We used the genes in a molecular signature to predict recurrence risk within 5 years after transurethral resection of bladder tumor. MATERIALS AND METHODS: Whole genome profiling was performed on 112 frozen nonmuscle invasive urothelial carcinoma specimens obtained at first presentation on Human WG-6 BeadChips (Illumina®). A genetic programming algorithm was applied to evolve classifier mathematical models for outcome prediction. Cross-validation based resampling and gene use frequencies were used to identify the most prognostic genes, which were combined into rules used in a voting algorithm to predict the sample target class. Key genes were validated by quantitative polymerase chain reaction. RESULTS: The classifier set included 21 genes that predicted recurrence. Quantitative polymerase chain reaction was done for these genes in a subset of 100 patients. A 5-gene combined rule incorporating a voting algorithm yielded 77% sensitivity and 85% specificity to predict recurrence in the training set, and 69% and 62%, respectively, in the test set. A singular 3-gene rule was constructed that predicted recurrence with 80% sensitivity and 90% specificity in the training set, and 71% and 67%, respectively, in the test set. CONCLUSIONS: Using primary nonmuscle invasive urothelial carcinoma from initial occurrences genetic programming identified transcripts in reproducible fashion, which were predictive of recurrence. These findings could potentially impact nonmuscle invasive urothelial carcinoma management.

Generation and external validation of a tumor-derived 5-gene prognostic signature for recurrence of lymph node-negative, invasive colorectal carcinoma.

BACKGROUND: One in 4 patients with lymph node-negative, invasive colorectal carcinoma (CRC) develops recurrent disease after undergoing curative surgery, and most die of advanced disease. Predicting which patients will develop a recurrence is a significantly growing, unmet medical need. METHODS: Archival formalin-fixed, paraffin-embedded (FFPE) primary adenocarcinoma tissues obtained at surgery were retrieved from 74 patients with CRC (15 with stage I disease and 59 with stage II disease) for Training/Test Sets. In addition, FFPE tissues were retrieved from 49 patients with stage I CRC and 215 patients with stage II colon cancer for an External Validation (EV) Set (n = 264) from 18 hospitals in 4 countries. No patients had received neoadjuvant/adjuvant therapy. Proprietary genetic programming analysis of expression profiles for 225 prespecified tumor genes was used to create a 36-month recurrence risk signature. RESULTS: Using reverse transcriptase-polymerase chain reaction, a 5-gene rule correctly classified 62 of 92 recurrent patients and 87 of 172 nonrecurrent patients in the EV Set (sensitivity, 0.67; specificity, 0.51). "High-risk" patients had a greater probability of 36-month recurrence (42%) than "low-risk" patients (26%; hazard ratio, 1.80; 95% confidence interval, 1.19-2.71; P = .007; Cox regression) independent of T-classification, the number of lymph nodes examined, histologic grade/subtype, anatomic location, age, sex, or race. The rule outperformed (P = .021) current National Comprehensive Cancer Network Guidelines (hazard ratio, 0.897). The same rule also differentiated the risk of recurrence (hazard ratio, 1.63; P = .031) in a subset of patients from the EV Set who had stage I/II colon cancer only (n = 251). CONCLUSIONS: To the authors' knowledge, the 5-gene rule (OncoDefender-CRC) is the first molecular prognostic that has been validated in both stage I CRC and stage II colon cancer. It outperforms standard clinicopathologic prognostic criteria and obviates the need to retrieve ≥12 lymph nodes for accurate prognostication. It identifies those patients most likely to develop recurrent disease within 3 years after curative surgery and, thus, those most likely to benefit from adjuvant treatment.

Applications of genetic programming in cancer research.

The theory of Darwinian evolution is the fundamental keystones of modern biology. Late in the last century, computer scientists began adapting its principles, in particular natural selection, to complex computational challenges, leading to the emergence of evolutionary algorithms. The conceptual model of selective pressure and recombination in evolutionary algorithms allow scientists to efficiently search high dimensional space for solutions to complex problems. In the last decade, genetic programming has been developed and extensively applied for analysis of molecular data to classify cancer subtypes and characterize the mechanisms of cancer pathogenesis and development. This article reviews current successes using genetic programming and discusses its potential impact in cancer research and treatment in the near future.

Feature selection and molecular classification of cancer using genetic programming.

Despite important advances in microarray-based molecular classification of tumors, its application in clinical settings remains formidable. This is in part due to the limitation of current analysis programs in discovering robust biomarkers and developing classifiers with a practical set of genes. Genetic programming (GP) is a type of machine learning technique that uses evolutionary algorithm to simulate natural selection as well as population dynamics, hence leading to simple and comprehensible classifiers. Here we applied GP to cancer expression profiling data to select feature genes and build molecular classifiers by mathematical integration of these genes. Analysis of thousands of GP classifiers generated for a prostate cancer data set revealed repetitive use of a set of highly discriminative feature genes, many of which are known to be disease associated. GP classifiers often comprise five or less genes and successfully predict cancer types and subtypes. More importantly, GP classifiers generated in one study are able to predict samples from an independent study, which may have used different microarray platforms. In addition, GP yielded classification accuracy better than or similar to conventional classification methods. Furthermore, the mathematical expression of GP classifiers provides insights into relationships between classifier genes. Taken together, our results demonstrate that GP may be valuable for generating effective classifiers containing a practical set of genes for diagnostic/prognostic cancer classification.

The use of genetic programming in the analysis of quantitative gene expression profiles for identification of nodal status in bladder cancer.

BACKGROUND: Previous studies on bladder cancer have shown nodal involvement to be an independent indicator of prognosis and survival. This study aimed at developing an objective method for detection of nodal metastasis from molecular profiles of primary urothelial carcinoma tissues. METHODS: The study included primary bladder tumor tissues from 60 patients across different stages and 5 control tissues of normal urothelium. The entire cohort was divided into training and validation sets comprised of node positive and node negative subjects. Quantitative expression profiling was performed for a panel of 70 genes using standardized competitive RT-PCR and the expression values of the training set samples were run through an iterative machine learning process called genetic programming that employed an N-fold cross validation technique to generate classifier rules of limited complexity. These were then used in a voting algorithm to classify the validation set samples into those associated with or without nodal metastasis. RESULTS: The generated classifier rules using 70 genes demonstrated 81% accuracy on the validation set when compared to the pathological nodal status. The rules showed a strong predilection for ICAM1, MAP2K6 and KDR resulting in gene expression motifs that cumulatively suggested a pattern ICAM1 > MAP2K6 > KDR for node positive cases. Additionally, the motifs showed CDK8 to be lower relative to ICAM1, and ANXA5 to be relatively high by itself in node positive tumors. Rules generated using only ICAM1, MAP2K6 and KDR were comparably robust, with a single representative rule producing an accuracy of 90% when used by itself on the validation set, suggesting a crucial role for these genes in nodal metastasis. CONCLUSION: Our study demonstrates the use of standardized quantitative gene expression values from primary bladder tumor tissues as inputs in a genetic programming system to generate classifier rules for determining the nodal status. Our method also suggests the involvement of ICAM1, MAP2K6, KDR, CDK8 and ANXA5 in unique mathematical combinations in the progression towards nodal positivity. Further studies are needed to identify more class-specific signatures and confirm the role of these genes in the evolution of nodal metastasis in bladder cancer.