Cox Proportional Hazards Assumption Violations in Breast Cancer Gene Expression Data

AJAS · 2022 Medicine

Doyoon Kim

Thumbnail supplied by the source for Cox Proportional Hazards Assumption Violations in Breast Cancer Gene Expression Data

Overview

Cox Proportional Hazards Assumption Violations in Breast Cancer Gene Expression Data. Background: Cancer is a disease driven by mutations in genes that control cell proliferation, cellular differentiation, and DNA damage repair. Prognostic biomarkers play an integral role in precision medicine by allowing clinicians to identify and treat patients with more aggressive tumours. A commonly used method for developing prognostic biomarkers is the Cox proportional hazards (Cox PH) regression model to estimate the change in hazard (risk) for a given mRNA abundance level for a particular gene. In this study, we tested the frequency of Cox PH assumptions, which genes violated the proportional hazards assumption, and assessed whether these violations were associated with specific biological features of cancer. We did so in seventeen breast cancer (2633 patients) and four Non-small-cell lung carcinoma (NSCLC) cohorts (433 patients), each with mRNA abundance and clinical outcome, comprising of 3076 total patients data available. Results: Neither the PH violation of the Cox model, nor the mean mRNA abundance level for a gene is correlated with the hazard ratio for patient survival in the Cox model. In a meta-analysis of the seventeen breast cancer datasets, cytogenetic bands, 5q14, 3p21, and 2p25 were identified as regions with enrichments in Cox PH violations with effect sizes of 1.4, 1.3, 1.3, and P-values of 0.212, 0.193, 0.231 respectively. However, the cytogenetic bands identified from the meta-analysis most likely have small effects on the accuracy of survival analysis biomarkers as the cytogenetic bands varied considerably across individual mRNA abundance datasets. Our results are consistent across breast cancer subtypes and in another common tumour type, lung cancer. Conclusions: While Cox PH violations are prevalent and exist, they are not strongly correlated with the biological features we analysed. PH assumption violations and mRNA abundance levels are weakly correlated with patient outcomes. Cytogenetic bands that are more likely to have Cox PH model violations exist but vary considerably across datasets, ultimately having little effect on Cox model violations. By studying the underlying assumptions, we provide insight into how we can improve Cox PH models and their reliability in clinical settings. Keywords Cox proportional hazards regression, breast cancer, survival analysis, biomarkers, meta-analysis

Video

If the video does not play, watch it on the original project page.

My Story

Hi everyone, I'm Doyoon. My journey in pursuing science research has been challenging, exciting, nerve-wracking, but overall, very rewarding! Throughout high school, I've had the opportunity to pursue lots of interesting projects and work with many talented, dedicated individuals and I'd like to share some parts of my story with you.

Growing up, I never considered myself to be much of a STEM kid. I always found myself gravitating towards the arts. I really loved music (still do!) and thought I'd end up pursuing music as a career. But in my freshman year of high school, I tried to branch out. I took AP Biology and really enjoyed learning about the complex biological systems in living things. I also learned some basic programming with Python.

There were two things that really motivated me to start pursuing science. 1) I attended a summer camp after 9th grade that specifically focused on the intersection music of technology. After working on a group project with some of my mates, I started to realize the limitless possibilities and applications of code in literally anything. 2) There was a heavy rise in news coverage related to automation and Artificial Intelligence in the workforce. I started reading about the many creative applications of AI in medicine and became interested in combining my love for Biology and my newfound interest in Computer Science.

I was interested in computer vision and image classification especially in diagnosing certain diseases or conditions. I started working on an independent project focused on diagnosing skin cancers using transfer learning deep learning models. There were definitely moments when I felt like I didn't know what I was doing and when I wasn't sure how to fix a certain bug or problem in my code. Nonetheless, I ended up writing a paper on this project and presenting at a few conferences and workshop events where I got to share my work with other researchers and scientists.

The experience from working on an independent project made me curious about what it was like to work in a lab setting with other scientists. I reached out to professors in my local area before I was able to start working with a cancer data science professor at UCLA. Here, I experienced a culture shock. My perceptions of an individual scientist working in their lab changed. By working with lab members on my code, editing my data plots, and learning a lot about statistics, biology, and ComSci topics, I realized the power of team science. I hope to continue exploring science through an interdisciplinary lens by approaching it from many different perspectives! Becoming an AJAS fellow is an amazing opportunity for me to learn more about how to do good science.