摘要
摘要
BACKGROUND: Topological data analysis has had many developments with potential application to precision oncology. The goal of this study is to determine if using geometry-based methods of gene-expression signatures of tumors can be used to both create a definitive classification of responders and non-responders, and to predict patient treatment response using out-of-sample testing methodology. METHODS: Gene-expression signatures of tumors, constructed from geometrical attributes of data, were created and tested in an open-sourced Pfizer clinical trial on avelumab plus axitinib in advanced renal cell carcinoma. This was a previously conducted and published Phase III randomized controlled trial as part of the JAVELIN Renal 101 project (2019). Multiple geometries of different properties based on data organization were made using gene-expression data (n = 726), which were then used to compile aggregate synthetic subgroups of response for every patient as an internal classification. Additional and separate out-of-sample testing of patients was conducted, using the remaining samples to build a geometrical scaffolding of response, and then creating a similar synthetic grouping then tested via the log-rank test. Results were compared against three published gene-expression signatures, one derived from this cohort and two externally. RESULTS: Geometrical gene-expression signatures were successfully used to create internal classification of responders to the intervention, with dramatically different Kaplan-Meier (KM) estimators based on responder assignment. Furthermore, out-of-sample prediction based on leave-one-out methodology was able to correctly predict this responder classification with 82.9% accuracy at median split of response, with area under the curve of 0.880 (95%CI 0.855-0.905). Gene-level analysis showed the signal to be a co-expressed program rather than individual transcripts. No gene retained an independent association after conditioning on its co-expressed neighbors, and two to five co-expressed modules exhausted transcriptomic information about classification (out-of-fold AUC 0.74-0.76), exceeding penalized regression on all 22,955 genes (0.73). Adjusted restricted mean survival time at 14 months was + 2.54 months in predicted responders versus + 0.44 in non-responders, an adjusted interaction of + 2.10 months (p = 0.004), which predicted responders accounting for 85.3% of the between-arm RMST gain. In unadjusted models, none of the published signatures reached a significant interaction (ratio of hazard ratios 0.67-0.82, all p > 0.06), against 0.55 (p = 0.005) for our classifier. Covariate adjusted Cox and RMST regression models gave consistent estimates at cutpoints below 10- 3, and 20,000 permutations tested supported the continuous interaction (p = 0.019; family-wise adjusted p = 0.042). CONCLUSIONS: Gene-expression based geometrical signatures can be used to create an internal reference classification of responders and non-responders in clinical trial data and are also highly accurate at predicting these labels out-of-sample, detecting treatment effect heterogeneity that established signatures did not. These methods can be used to identify target biomarkers of response, which may support enrichment or stratification designs.
KM 提取结果
0 个 Figure · 0 个 KM 面板