Growing tumours are evolving systems where cells continuously gain mutations that influence their survival and the whole tumour's future progression. The question is, can we decode these processes from the mutations detected in a single cancer sample? In the Evolution and Cancer Lab at Barts Cancer Centre, we have a long history of asking such questions. Over the years, we used a toolset borrowed from population genetics and stochastic modelling to figure out how the mutation spectrum looks under neutral evolution (when "all cells are equal", Williams et al., 2016) and positive selections (when a few cells "are more equal than others", Williams at el., 2018). (Turns out, there are theoretical patterns, which sparked a stimulating debate in the field since then.) In this work, we set out to study the missing piece: what does the mutation landscape look like under negative selection (when many cells are "less equal than others")?
In particular, we were interested in neoantigen-associated mutations, mutations that help the immune system recognise and eliminate cancer cells. We used a mathematical model to simulate the growth of a tumour, in which the cells (i) gain neutral and antigenic mutations upon every cell division, (ii) are killed by the immune system according to their total antigenicity and (iii) can gain immune escape. Using the perks of simulations (no ethical or practical limitations) we could track cancer growth over time to establish the hallmarks of negative selection.
We observed that the proportion of the tumour carrying an antigenic mutation declined during tumour growth - since cell with such mutations had a higher probability to be killed and got outgrown by their wild-type (neutral) counter-parts. This is exactly the process of immuno-editing: consistently pruning cells that would tip off the immune system. In the final tumour population, this led to a "gap" in the frequencies of antigenic mutations: while neutral mutations were distributed evenly, negatively selected mutations were more concentrated on the rare end of the spectrum. This relationship showed up clearly when plotting the cumulative number of mutations against their inverse frequency in the tumour population, creating a characteristic "dip" in the graph. (Which was, reassuringly, in agreement with population genetics results, e.g. Cvijovic et al., 2016.)
But is this characteristic depletion a common hallmark of cancer genomes then? We found that there were two major challenges in detecting this pattern in real-life - making negative selection a rather elusive phenomenon.
On the one hand, the depletion of negatively selected mutations meant they were typically present in only a small number of cells, something that became more pronounced with higher selection (stronger immune system). Taking sequencing depth into account - where only mutations present in >0.1% of the tumour are detected - we found that only a handful of true neoantigens were picked up. This caused (i) the overall mutation spectrum to look neutral (as >99% of detected mutations were neutral), and (ii) a lack of statistical power to evaluate the difference between neutral and antigenic mutation spectra. Therefore our power to detect negative selection declined as selection strength increased - somewhat counter-intuitively leaving medium-selection tumours as the best candidates to capture selection. (As a side note, we would not have quantified this relationship without comments from Reviewer 2 - this paper matured dramatically in the review process.)
On the other hand, we found that most tumours carrying/accumulating a substantial neoantigen burden (i.e. tumours that are hyper-mutated) would be eliminated rapidly, and could only grow to a detectable size if they developed immune escape. Having circumvented the pressure from the immune system meant that in these tumours all mutations became neutral again (so no selection signal). In practice, being able to sample only tumours above a detection limit, we have an observation bias for immune escaped cancers where negative selection is not active anymore - again lowering our chance of detecting this selection.
So what is the landscape of real cancer samples like? We analysed colorectal (CRC), stomach (STAD) and endometrial (UCEC) cancers from The Cancer Genome Atlas and found that immune escape was an extremely wide-spread phenomenon, especially amongst cancers with a high mutation load due to mismatch repair (MMR) or polymerase epsilon (POLE) deficiency. Following our model predictions, we evaluated negative selection on a select subset of these cancers: without immune escape, not hyper-mutated (microsatellite stable, MSS) and showing an intermediate level of T-cell infiltration that suggested medium-level immune selection. We pulled all cancers into one mutation cohort, and this way we could capture sufficient mutations to see a deviation between the total mutation spectrum and neoantigens-associated mutations, confirming that indeed these mutations are experiencing negative selection.
Overall, we found that there is a hallmark of negative selection: the depletion of high-frequency antigenic mutations. However, most cancers lack this hallmark which highlights the high prevalence of immune escape. This might be good news as some immune escaped tumours can be targeted with immunotherapy - but for that we need a better picture of the full history of negative selection during tumour growth. We explore further avenues in our paper, but at the end of the day there is still much to learn about the tumour-immune co-evolution during cancer development.