Does immune response change in different parts of lung tumor? To what extent the spatial configurations of immune and stromal cells influence the evolution of immune escape? 

As the computational pathology team in the TRACERx program and in partnership with expert pathologists, oncologists, and bioinformaticians, we use technological advances in deep learning and computer vision to address these questions. 

In the process, we developed a generalizable artificial intelligence (AI) framework which produces a spatial map, where all cancer, lymphocyte and stromal cells can be recognized. The AI model was trained to follow the way thoracic pathologists identified these cells, using 21,009 single-cell annotations. Due to the high throughput nature of this technology, we analyzed 5 billion cells using ICR's high-performance computers. 

With this framework, we were able to spatially profile immune infiltration from histology slides collected from different parts of lung tumors. Surprisingly, different tumor regions can be dramatically different in the amount of immune cells, even within the same patient. We differentiate them in our analysis as "immune hot" or "immune cold" regions by using the percentage of all cells that were lymphocytes in each histology slide. 

Owing to the strong correlation between our method and i) RNA-seq-based immune profiling (from adjacent region-specific tissue, Rosenthal et al. 2019) and ii) pathology tumor-infiltrating lymphocyte scoring (using the same histology sections, Hendry et al. 2017), even simple immune cell quantification from routine histology sections, though perhaps not "deep" in its profiling, can quickly and accurately measure the immune response of lung tumors. 

However, a key AI pathology challenge is how to validate the trained models. 

When we talk about AI for more frequent computer vision tasks (e.g. face detection or a cat versus dog problem), we are looking at significantly larger ground truth datasets (often millions of training events). While virtually anyone can -with high confidence- label a cat or a dog, this is not as easy for experts when it comes to distinguishing different tumor microenvironment cells in heterogeneous lung samples. 

Pathologists often express discomfort if I randomly selected an image tile and asked them to annotate each and every cell. Furthermore, this challenge is amplified when we have to consider that only experts with busy clinic schedules can perform such a sophisticated annotation job. So with a limited ground truth data, what is the real "test"? 

Our study draws a roadmap to address this challenge in two steps. 

• Large scale external validation: 

Trained AI models as well as survival tests from TRACERx were independently validated in a large retrospective cohort of 970 lung cancer patients, the LATTICe-A study (Moore et al. 2019). With 4,324 tumor blocks, this is the largest multisample digital pathology application to-date. 

We can call this an "end-to-end" AI model validation. LATTICe-A samples were scanned well before we even began to analyze the TRACERx cohort.

• Single-cell "biological" validation: 

To overcome the limitation in single-cell ground truth data, we virtually integrated IHC and H&E images generated from the same slides. Cancer, lymphocyte and stromal cell identification was validated using specific markers of cell types: TTF1, CD45 and SMA staining, respectively. 

It's worth noting that with this experiment, we generated 137,603 cells in a single run for three antibodies. This would have consumed a huge amount of time for an expert to manually annotate. 

The crux of the matter lies with careful AI training, we relied on diverse histology samples covering as many patients as possible from the TRACERx cohort. Our AI models are trained to capture cell biology, just like a trainee pathologist.  

Taking these steps together, this study provided evidence for the AI pathology system to generalize from one patient cohort to another, which was processed and digitized in a different center. 

I hope with this effort, we are one step closer to seeing computational pathology as a frequent, assistive tool in the clinic. 

With this AI pathology immune classification we learned that more significant than any other immune feature tested, simply just the number of immune cold regions, is the best predictor of cancer relapse. Specifically, tumors with more than one immune cold region had a higher risk of relapse, independently of tumor size, stage and number of samples per patient. This finding was first discovered in the TRACERx cohort and it was validated in an independent multisample cohort of 970 patients with lung adenocarcinoma.

Our study strongly suggests that it is not enough to measure immune response using a single sample to represent the whole tumor, a finding that was only possible because of the multiregion study design in TRACERx. 

Even within a tumor 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. But, why? 

To answer this question, we investigated the pairwise cold-cold versus hot-hot genetics within the same tumors. We wanted to know what is so special about a tumor that has more than one immune cold region.

Within the same lung tumours, geographically dispersed immune hot regions share early, clonal cancer mutations. In contrast, tumors which harbour an increased number of immune cold regions tend to share late, subclonal cancer mutations, and therefore are more lethal to treat. To visualize this differential evolution history of tumor regions, we mapped our AI pathology immune classification on the TRACERx phylogenetic data (Jamal-Hanjani et al. 2017): 

There could be a plausible explanation here. 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' subclones might have originally emerged in immune hot environments, but then through genetic events acquired, the region is turned cold.

What we are seeing here is a “more recent” common ancestor in cold regions, which may have been selected due to its ability to evade the immune system. We speculate that by identifying these lethal immune-escaping subclones, new drivers of immune evasion may be elucidated.

AI enabled us to measure significant geographical variability, seen as “hills and valleys” of immune response, echoing differential clonal structure of cancer evolution within the same tumors. 

We are embarking on the much needed "Galapagos" geography to better understand the evolutionary interplay of cancer and the immune microenvironment (McGranahan & Swanton, 2017). 

When we further dug into in the spatial make-up of the immune hot and cold phenotypes, we found two interesting spatial patterns. In lung adenocarcinoma but not lung squamous cell carcinoma, both patterns are related to known genetic alterations relevant to immune surveillance: 

• Compared with immune hot regions, immune cold regions tend to have a more irregular geometric interface between cancer and stromal cells, as measured by fractal dimension. 

This new measure estimates physical constraints against T cell ingress and it was related to dysfunction in antigen presentation through loss of heterozygosity at the human leukocyte antigen locus (HLA LOH). A significantly higher fractal dimension was found in tumor regions with intact HLA alleles compared with regions harboring HLA LOH. 

• Using unsupervised clustering, we found a specific spatial subset within immune cells, restricted to "adjacent-to-tumor" compartment that was associated to clonal neoantigens. 

To summarize, the collective and independent validation of our AI models is a step toward translating computational pathology. Histology-deep-learning framework infers key ecological processes shaping clinical phenotypes and genetic clonal diversity.

TRACERx is a pioneering cancer evolution study. Much of what we are capable to analyze today is due to large collaborative effort and planning from study founders and the consortium 4-5 years ago.