TOOL TO PREDICT CANCER SURVIVAL

Estimating how long a cancer patient may survive is always tricky: a new statistical method appears to provide accurate information, although it will not be available for a year or longer.

The new method, known as SURVIV (Survival analysis of mRNA Isoform Variation), models the measurement uncertainty of mRNA isoform ratio in RNA-sequencing data to predict patient survival time.

In simulation studies, it consistently outperformed conventional Cox regression survival analysis, especially for datasets with modest sequencing depth, researchers report in a study published online June 9 in Nature Communications.

These simulation tests were carried out for six different cancer types: invasive ductal carcinoma (IDC), glioblastoma multiforme, lower-grade glioma, renal clear cell carcinoma, lung squamous cell carcinoma, and ovarian serous cystadenocarcinoma.

"Using biomedical big data that encompass the molecular and clinical profiles of cancer, we can identify novel biomarkers that guide cancer prognosis and treatment," said lead author Yi Xing, MD, of the Department of Microbiology, Immunology and Molecular Genetics at the University of California, Los Angeles.

"More specifically, we found that isoform-based predictions work consistently better than the conventional gene-based predictions in predicting survival time," he told Medscape Medical News.

"We hope [we] can potentially provide a powerful 'precision medicine' tool that can be used in the clinic to guide cancer treatment," he said in an interview. He estimated that this could take 1 to 3 years.

Simulation Tests

Dr Xing and collegues spent more than 2 years developing the SURVIV algorithm. They used histologic samples from a total of 2684 cancer patients.

The first simulation test was carried out for 682 patients with IDC from the Cancer Genome Atlas (TCGA) breast cancer RNA-sequencing dataset.

To compare the performance of SURVIV with a conventional Cox regression survival analysis using the point estimates of exon-inclusion levels, the researchers designed a simulated data set of 600 individuals that mimicked the parameters of the TCGA IDC breast cancer data.

In each simulation, they simulated 20,000 alternative exons to correspond to the number of exon-skipping events in the TCGA IDC data. A total of 90% of the exons were from the null hypothesis — that the exons were not associated with survival time. The remaining 10% of the exons were from the alternative hypothesis — that exons were associated with survival time.

The simulated data were then used in two settings: in the setting without censoring, the death and survival time of each patient was known; in the setting with censoring, the death and survival time was not known. The latter included patients who were still alive.

The researchers mimicked the censoring rate of the TCGA IDC dataset by assuming that 85% of the patients were still alive at the end of the study. "In both settings and with different depths of RNA-sequencing coverage, SURVIV consistently outperformed Cox regression in the true-positive rate at the same false-positive rate of 5%," Dr Xing and colleagues report. "As expected, we observed a more significant improvement in SURVIV over Cox regression when the RNA-sequencing read coverage was low."

When the SURVIV analysis was carried out in five additional cancer types in TCGA, lower-grade glioma had the strongest survival-associated alternative splicing signals with 660 significant exon-skipping events. "Strikingly, regardless of the number of significant events, alternative splicing-based survival predictors outperformed gene expression–based survival predictors across all cancer types consistent with our initial observation on the IDC data set," the researchers write.

Evaluation of the different types of predictive models showed that a model incorporating clinical data, gene expression, and alternative splicing data performed significantly better than any one type of predictor alone.

Now the team is applying SURVIV to much larger datasets across many more types of cancers to develop more reliable isoform-based predictors of patient survival. They are hoping to discover isoforms that are consistently associated with survival in a "pan-cancer" analysis across multiple cancer types, Dr Xing explained. This method could be extended to predict other types of patient outcomes, such as response to specific therapies.