Image-Omics The Yuan Lab Publications Team Projects Contact
TRACERx

TRACERx


Computational pathology and AI to study lung cancer evolution

    TRAcking Cancer Evolution through treatment (Rx) (TRACERx) lung study is a pioneering study of lung cancer evolution. Experts from a wide range of disciplines work closely to integrate large-scale clinical, histopathological and genomic data from patients with lung cancer. The study aims to define how cancer clonal heterogeneity affects the risk of recurrence and survival, and how cancer subclones compete, adapt and evolve from diagnosis to relapse. Thus far, data on the first set of 100 patients have been generated to answer intriguing scientific questions as a result of this consorted effort (Nature special collection). We are part of this landmark study funded by Cancer Research UK to deliver technological advances in computer vision and digital pathology to lung cancer research.

Cancer’s lethal ability to evolve and become resistant to treatment is one of the biggest challenges facing cancer researchers and doctors today. We apply machine learning and computer vision innovation to pathology to tackle this challenge. Our aim is to better understand cancer’s evolution by studying its environment.

 

Generalisable, accurate, and explainable AI for cancer research

In our first publication on TRACERx and digital pathology in Nature Medicine (AbdulJabbar et al. 2020), we presented an AI system that spatially profiles various cell types from lung histology slides and predicts lung cancer relapse. Our study provided evidence for the AI pathology system to generalize from one patient cohort to another, by validating our AI models on 4,324 tumour blocks from 970 patients in an independent cohort. To our knowledge, this is the largest multi-sample digital pathology application to-date.

Validating the accuracy of the AI system is a challenge. We used five orthogonal data types to "challenge" our AI system, including DNA next-generation sequencing (DNA-seq), RNA-seq, pathological immune scoring. We also developed a new way to accelerate the process of collecting cell-type specific data for testing, by virtually combining different types of histology samples stained on the same section, generating 137,603 cells in one experiment. This amount of cells would have taken our pathologists >100 hours to annotate.

We believe that reproducible biomarkers are based on biology, and to deliver new biology-based biomarkers, it is essential that we work closely with our collaborating pathologists. Therefore, instead of directly training AI against cancer outcome, we take a different approach. We co-develop our AI system with pathologists, transferring their expert knowledge about the immune system. This allows us to generate explainable and biologically meaningful features that may be used as biomarkers.

 

A new feature of immune response to predict risk of cancer relapse

A striking observation in this study is that the amount of immune cells can sometimes change dramatically within the same tumour. In these cases, areas of tumours with few immune cells co-exist with areas packed with immune cells. Remarkably, these areas of tumours with few immune cells, termed "immune-cold" regions, appear to have high clinical relevance. The number of immune-cold regions is the most predictive feature of risk of cancer relapse in our study. Many previous studies have shown that measures of immune response may have value in predicting cancer outcome. However, this study shows that it is not enough to measure immune response using a single sample to represent the whole tumour. High number of immune-cold regions in a tumour predicts high risk of lung cancer relapse, better than the number of "immune-hot" regions, or the average amount of immune cells in the tumour.

This finding is only possible because of the multi-region sample collection study design in TRACERx.

Differential cancer evolution echoes immune variability

Thus, even within a tumour that has on average increased immune cells, if it contains regions classified as immune cold, prognosis appears to be associated with the number of cold regions. The question is why? Analysis of genetic data generated from the same samples in TRACERx offered some answers to this question.

We found different evolution patterns of lung cancer cells in the same tumour, this differential evolution is reflected in the variation of immune activity, as seen by our AI system. Cancer cells lived in immune cold regions tend to evolve later than cancer cells lived in immune hot regions, perhaps after developing an ability to escape from immune cell predation. Cold regions tend to share a “more recent” common ancestor than hot regions, which may have been selected due to its ability to evade the immune system.

We speculate that by identifying the lethal immune-escaping subclones, new drivers of immune evasion may be elucidated. This can be a vital step for future targeted drug development.

Similar to how Charles Darwin studied finches in the "Galapagos" islands, our AI-histology approach in TRACERx explores each individual tumour’s environment to better understand how it grows and, further, how to stop it from developing treatment resistance. There is much to be discovered in this project, as we move on to the next stage, analysing more patient data collected by the TRACERx consortium.

Paper | GitHub | Blog | TRACERx Nature special collection | Postdoc position available