Our next discussant is Dr. Mattson, who's the recipient of the Columbia Softball Pediatric Award. The title of his talk is, In silico identification of neo-antigens in high-grade pediatric brain tumor cohort utilizing next-generation sequencing, pilot study of a discovery pipeline for immune therapy targets. All right, thank you everybody, and thank you for the scientific committee for this chance to talk. My name is Peter Mattson. I'm a resident at the University of Pennsylvania. I'll be giving a description of our work on identifying neo-antigens in high-grade pediatric brain tumors utilizing next-generation sequencing. So the concept of a neo-antigen has become important in recent years in cancer biology. In a tumor cell, point mutations, insertions, deletions, and fusions can introduce non-synonymous sequences into the tumor genome. If these occur in translated regions, they will then be converted into novel proteins not present in the rest of the body. Fragments of these novel proteins, like all proteins in a cell, will be processed and presented to the immune system via binding to class I and class II HLA molecules. These novel peptides bound to patient-specific HLA molecules are considered neo-antigens. Since they are not found in normal tissues and they are expressed on the outside of tumor cells, they make a nice, attractive candidate for targeted immunotherapies. So recently, it has become possible to predict with reliability the neo-antigens that will be generated by a patient's tumor cell in silico. Whole genome sequencing or whole exome sequencing of a tumor sample is used to identify the mutations that are present. If they occur in the translated region, the resultant peptide sequence is then used to create a tiled series of smaller peptides around that mutation. These novel peptides are then fed through algorithms using machine-learning techniques to identify and predict how they will bind to those HLA molecules. There's incorporation of whole genome sequencing at that point as well for the HLA prediction. As a result of that, you get a nice long list of candidate neo-antigens, and they're typically filtered by their binding strength. We shoot for neo-antigens with a binding strength less than 500 nanomolar. We also can include RNA-seq data to get expression results to narrow things down based on just those neo-antigens that are expressed. This workflow for neo-antidiscovery has the potential to perform a number of various cancer immunotherapy techniques. In one instance, a generation of a panel of patient-specific neo-antigens can be used to generate a personalized vaccine. Vaccine administration, it stimulates a tumor-specific immune response that has been shown in some small trials in adult cancers to be efficacious and safe. These vaccines can either be completely personalized or they can include shared neo-antigens in sort of an off-the-shelf vaccine standard. Predicted neo-antigens could also be used to generate adoptive immunotherapies such as TCRs or CARs. Usually these are for shared antigens. And then you can also use this workflow to predict response to certain immunotherapies such as checkpoint inhibitors. So as a group, pediatric high-brain tumors such as DIPG, high-grade glioma, medulloblastoma, and ATRT are, you know, histologically diverse, but they do share certain similarities, one being their poor prognosis and lack of efficacious therapies. Also there's some common underlying genomic and epigenomic signatures in these tumors that may lead to common treatment strategies. For these reasons, we set out to define the neo-antigen landscape in this group as a whole. So our work was enabled by two consortiums, the CBTTC and PNOC, where they've been able to assemble a large collection of pediatric brain tumors and do whole genome sequencing and RNA-seq on that group. We used, like I said, sort of an umbrella approach to our patient selection to look for cross-sectional diversity and to capture tumors that potentially have higher mutational burdens. Our bioinformatic tools included OptiType for HLA prediction, as well as NetMHC, which was our cloud-based neural network that we used for binding prediction. Work was all performed on an open-source, cloud-based genomic data platform that we've developed called Covatica. So our patient cohort, we were able to compile. These all have complete whole genome sequencing and RNA-seq. We had 17 patients with high-grade glioma, 15 with DIPG, 12 with beduglastoma, and two with ATRT. So we first defined the mutational burden in these tumors, and across the categories, total mutations, which included point mutations, insertions, and deletions, ranged from about 2,000 to over 10,000. The average mutation, which in each group per megabase pair, was about just over one for all the groups. This is lower than many adult cancers, like melanoma or lung cancer, but it's still high enough to support this neoantigen strategy. So we then process these samples through our algorithm for prediction of the class 1 neoantigens. So this volcano plot on the left shows the number of mutations in each sample that generate neoantigens. On the blue is the data before we take into account expression, and gray is after taking into account expression. And on the right side, you can see the distribution of the number of predicted neoantigens per sample, again, with and without the expression data. And within the groups, there's about a range of 10 to 15 on average mutations that lead to neoantigens, resulting in about 30 to 80 neoantigens per sample on average. And then after an expression data is taken into account, that goes down to about maybe 7 to 30. So then we did the same thing for our class 2 prediction algorithm. Again, on the left side there, you see the distribution of the mutations that are leading to neoantigens. And on the right side, you have the distribution and the number of neoantigens that we found in each sample. You can see the scale on the number of predicted neoantigens on this cohort is much larger. It has to do with the fact there's about…there's twice as many HLA alleles to bind with for these peptides, so that increases the number significantly. So taking these together, you get the class 1 and the class 2 data. Across all samples, there's a really sufficient number of neoantigens that we're finding here to support investigation into this type of work. So then when we took a little bit closer look at the class 1 data, we wanted to look for any shared neoantigens across the cohort. So we identified actually 19 different genes that contributed identical neoantigens in at least two or more subjects. We considered these shared neoantigens. And unexpectedly…not unexpectedly, you know, the most common one that came out was the product of the histone mutation found in DIPGs and some high-grade gliomas, the H3K27M mutation. But you know, there were a large distribution of other genes as well. The one thing that we found that was quite interesting is that we found identical or shared neoantigens between different histologies, and this wouldn't have been found if we had taken a much more narrow approach to patient selection. And we had an overall 60 different novel epitopes that resulted in shared neoantigens, and two-thirds of patients had shared neoantigens between the groups. So in conclusion, you know, we described some immunogenomic techniques that utilize next-generation sequencing to enable novel discovery of targets, neoantigen targets in immunotherapy. We found that the mutational burden and the number of neoantigens across these tumor types were high enough to suggest that we should be looking into this as therapeutic strategy. And shared neoantigens actually seem to occur more often in this group than in the adult cancers. This may have to do with sort of the developmental influences that occur in the way that these tumors develop, and they can lead down certain conserved mutational signatures. And the shared epitopes we're finding are beyond the H3K27M mutant, and they're across tumor types. This has significant implications for when we're thinking about designing vaccine trials. So for future directions, we're refining our pipeline further to better include the RNA-seq data and working to better refine that Class II data. And then we're also making a pipeline for fusion product identification and neoantigen prediction from that. We're expanding our cohort to include all pediatric brain tumors within CVTTC, and we're working to correlate all this data with survival and treatment response. And then ultimately, we're working on clinical trial design based on this information, and we're hoping this can lead to better clinical decision-making as well. So of course, I work with a big group at CHOP. I just have to thank them, especially Jay Storm and Adam Resnick, our mentors. Thank you.

Dr. Peter Mattson

neo-antigens

high-grade pediatric brain tumors

next-generation sequencing

immunotherapies