Preliminary development of a predictive signature for response to palliative chemotherapy in gastro-oesophageal adenocarcinoma using multilinear singular value decomposition
Session type: Poster / e-Poster / Silent Theatre session
Multi-centre biomarker study UKCRN12435 is analysing multiple circulating biomarkers including plasma IL6, IL8, IL10, TNFa, and IFNg longitudinally in advanced gastro-oesophageal adenocarcinoma patients receiving palliative chemotherapy. We previously reported baseline and cycle 2 (C1/C2) survival associations for IL6, IL8, and IL10 (NCRI 2017/2018). No survival-associated cytokines relate to objective response at either time. To investigate a combinatorial signature, we analysed 71 patients with full response, clinicopathology, and C1/C2 cytokine data using multilinear singular-value decomposition.
Elements Tijk of third-order tensor T represent patient i, cytokine j, at time k. Decomposition yields:
Tijk = Sabc . Uia . VTbj . WTck
Summation is implied over a, b, c; S is a rank-(10, 5, 2) tensor. Columns of U can be regarded as "building-blocks" for observed all-patient cytokine measurements. Sequential forward feature-selection using fitting-deviance of a logistic model of response with candidate covariates U, TNM-stage, gender and age, selected three columns of U, gender, T-stage, and N-stage.
2/3 columns of U and all clinical covariates were significant (p<0.05). Classification accuracy was 76% (responder/non-responder) with AUC=0.835. Sensitivity and specificity were 83% and 62.5%. Neglecting U, only N-stage was significant (accuracy 69%; AUC=0.717; sensitivity 78.7%; specificity 50%). Likelihood ratio test confirmed significance of cytokine basis terms (p=0.0027). A comparator using forward-selected raw cytokine measurements performed more poorly.
Raw cytokine levels are poor response predictors. Though preliminary, C1/C2 tensor analysis improved accuracy modestly but significantly over clinicopathology-only; specificity was most improved. UKCRN12435 is recruiting ~400 patients, comprising 2 cohorts (biomarker discovery and validation cohorts). We have fully profiled the discovery cohort baseline (183 patients), but only 114 patients at C2. Longitudinal studies are continuing and we are also investigating additional immuno-oncology markers. A larger training set, additional markers, and/or cycle 3 data may improve classification. We await results from cohort 2 to confirm model validity.