Prediction of Pathologic Complete Response by Gene Expression Profiling in Esopha
Principal Investigator: Jaffer Ajani
Affiliation: The University of Texas
Abstract: Our objective is to develop a PCR-based ~10-gene signature, through gene expression analyses, that can predict all three subtypes of pathologic responses (with high accuracy) following chemoradiation therapy in patients with esophageal cancer who undergo chemoradiation followed by surgery (Tri-modality [TM] therapy). The three pathologic subtypes are: pathologic complete response (pathCR), partial response, and extreme chemoradiation-resistance (exCRTR). One can conceive a therapeutic approach suited for each outcome (e.g., avoid chemoradiation in patients whose cancer has an exCRTR). Today however, there are no tools to optimize therapy for these outcomes since we cannot predict them before therapy. A predictive signature that has a high level (=80%) of specificity and a reasonable level of sensitivity (=45%) would be an advance. Our hypothesis is that a practical molecular signature can be established through gene expression profiling to predict three subgroups prior to TM therapy. In our 19-patient gene expression profiling study, the unsupervised hierarchical cluster analysis segregated cancers into two subtypes. Five of 6 pathCR patients clustered in subtype I and one pathCR patient clustered in subtype II. We discovered that Sonic Hedgehog and NF-kB-related genes appear to mediate chemoradiation-resistance. We were able to independently validate this. In a gene expression analysis of 47 TM patients (Specific Aim 0), we used 17 genes (10% false-discovery rate) to construct a multivariate model to predict response. For each gene g, we first computed the residuals Rg,i from a linear model of the form , where Yg,i is the expression of gene g in sample i, t(i) is the subtype of sample i, and Sg,t(i) is the mean expression of gene g in samples of that subtype. We then used the residuals as predictors in an ordinal regression model to predict the outcome categories. We used the Akaike Information Criterion (AIC) to remove unnecessary variables from the model. The final model involved 7 genes: RiskScore=1.59 TMEM46 + 0.68 THBS1 -1.52 LOC442578 - 2.14 SRM 1.16 CHST4 + 0.83 DES + 1.14 SDS, with a cutoff between pathCR and partial response at -1.56 and a cutoff between partial response and exCRTR at 3.72. Four of these seven genes are related to Sonic Hedgehog pathway and 2 are NF-kB targets. In this proposal, data from 120 TM patients to be analyzed through a funded grant (R21CA127612) will be added to a new cohort of 120 TM patients (Specific Aim 1) to establish a large (n=240) training (discovery) set. We will identify best performing ~100 genes through microfluidic card technology. Specific Aim 2 will validate ~100 best genes and refine the model to select ~10 best performing genes for predicting three outcomes. Specific Aim 3 will prospectively validate the ~10-gene signature. A continuous "risk score" for the outcome will be computed. Specificity and sensitivity will be determined by generating receiver-operating (ROC) curves for optimizing the prediction boundaries. PUBLIC HEALTH RELEVANCE: This proposal is an early attempt to individualize therapy based on molecular biology for patients with esophageal cancer. Our goal is to pave the way for a strategy in the future that will allow administration of effective therapy, improve safety, and preserve the esophagus in some patients.
Funding Period: -------------------- - --------------------
more information: NIH RePORT