Carlos Traynor (1,2), Tarjinder Sahota (2), Neil Evans (1), Helen Tomkinson (2), and Michael Chappell (1).
1 University of Warwick, Coventry, UK. 2 AstraZeneca, Cambridge, UK
Objectives: Lung cancer is the second most common cancer and the leading cause of cancer related death worldwide. In the UK, 46388 new cases were diagnosed and 35620 people died of lung cancer in 2016 [1]. Recently, new tailored therapies directed at the underlying molecular cause are being added to the battery of available anti-cancer drugs. Next-generation sequencing (NGS) is a technique based on modified DNA microarrays, a high-throughput technology that measures thousands of gene expressions in a single experiment in different individuals. The aim of this project is to use non-small cell lung cancer (NSCLC) patient data from a bespoke database to model and predict disease progression (DP) and overall survival (OS) and to further evaluate the association between biomarkers (related to NGS) and clinical outcome. To this end, we have developed mctte an open-source R package, freely available on GitHub [2], which allows fitting stochastic state-space models applied to time-to-event data. It builds upon recent developments in Bayesian software and uses Stan [3] for parameter estimation.
Methods: The data obtained from The Cancer Genome Atlas (TCGA) [4] comprise the clinical outcomes for 495 NSCLC patients, overall survival or time-to-death and disease progression or time-to-recurrence. Disease progression is evaluated by an increase in cancerous lesions or the appearance of metastasis. The data also include clinical covariates and 20431 NGS measurement. To model the course of NSCLC we propose an absorbing Markov Chain with disease progression as a transient state and death as an absorbing state. The model is a proportional Markov time-to-event model that accommodates censoring, allows covariate effects and assumes that baseline hazards of stable to deceased, and recurred to deceased are related through a proportional hazard ratio on the transition.
Results: Age has a higher risk for non-specific cancer death, while stage has a
higher risk for related cancer events. Smoke might be easily misinterpreted
while smoking people have a higher risk of contract lung cancer, the clinical outcome may differ depending on cancer subtype. The results for NGS are more informative to prognostic biomarkers: MYL7, RGL3 and MPRIP are related to cytoskeleton remodelling and actin binding, but with opposite hazard ratios in the present study. PDCHA3 and FMN1 are related to cadherin and cell adhesion. VAV2 is related to EGFR signalling.
Conclusions: The novelty in this study includes the development of a mechanistic model of disease progression from clinical, genomic and survival data, for lung squamous cell lung carcinoma. This study showed that the hazards of lung cancer-specific events are associated with genomic variables related to actionable mutations that are expected to predict the effect of therapy. The design of a forecasting model of disease progression combined with treatment effects may be used to select patients, thereby, advancing towards personalised therapies.
References:
[1] Cancer Research UK https://www.cancerresearchuk.org
[2] https://github.com/csetraynor/mctte
[3] Stan Development Team. 2018. The Stan Core Library, Version 2.18.0.
[4] Gao et al. Sci. Signal. 2013 & Cerami et al. Cancer Discov. 2012
Reference: PAGE 28 (2019) Abstr 8997 [www.page-meeting.org/?abstract=8997]
Poster: Methodology - New Modelling Approaches