NEW YORK (GenomeWeb) – A team led by researchers at pharma firm Gritstone Oncology has developed a computational model for predicting cancer neoantigens.
Detailed in a paper published this week in Nature Biotechnology, the model, named EDGE, improves upon existing methods of predicting which neoantigens will be presented to the immune system by cancer patients' human leukocyte antigen system, said Roman Yelensky, executive vice president and chief technology officer at Gritstone and senior author on the study.
Yelensky said that better neoantigen prediction could aid in the development of more effective immunotherapies and noted that Gritstone is currently using the EDGE model in clinical trials testing personalized cancer vaccines. The company has also partnered with Cambridge, Massachusetts-based drug developer Bluebird Bio to use the model for target identification and is pursuing additional partnerships with researchers in pharma and academia, Yelensky said.
Expressed on the surface of most cells, the human leukocyte antigen (HLA) complex presents peptides produced within the cell to the body's immune system. In the case of, for instance, a viral or bacterial infection, the HLA complex will present peptides processed from these infectious agents, which triggers an immune response, stimulating T-cells to attack cells displaying these foreign peptides.
Much research in cancer immunotherapy has focused on identifying cancer-specific peptides, or neoantigens, that HLA complexes are likely to present to the immune system, the idea being that were these neoantigens identified, the immune response to them could be enhanced via, for instance, a vaccine.
A major difficulty, however, has been identifying which neoantigens will be produced and presented by HLA complexes at sufficient levels.
Traditionally, Yelensky said, researchers have done this by screening neoantigen peptides against HLA molecules to determine which pairs have strong binding affinities, the thinking being that neoantigens with stronger binding affinities are more likely to be presented by an HLA complex. Using this peptide-HLA binding data, researchers have developed machine-learning tools that can be used to predict peptide-HLA binding based on their genomic sequences.
"The issue is that while that kind of approach can work quite well for predicting binding affinities, what it doesn't do is capture what happens before HLA-peptide binding," he said. "So you might predict that a given neoantigen would bind well to a patient's HLA complex, but that gene might never be expressed at a level high enough to produce enough [neoantigen] peptides to even get to that [HLA-peptide binding] step."
That has led to overprediction of neoantigens, Yelensky said. According to the Nature Biotechnology study, less than five percent of peptides predicted by such tools were actually found on the surface of patient cells.
To address this problem, Yelensky and his colleagues used mass spectrometry datasets to identify the neoantigen peptides actually present in patient samples. They analyzed 74 cancer patients, identifying an average of 3,705 peptides per sample. They also collected transcriptomic data on these samples. Using this along with other previously generated HLA datasets, they developed a neural network model to predict neoantigens likely to be presented by HLA complexes.
"So, our model, instead of being trained on HLA-peptide binding data, was actually trained on peptides that were presented by HLA in human tumor specimens," Yelensky said.
As the authors noted, this approach was not entirely novel, as several recent HLA-peptide binding models have incorporated neoantigen data generated by mass spec analysis of tumor cell lines. However, in addition to using neoantigen data collected from actual cancer patient samples, the EDGE model also incorporated transcriptomic data, which Yelensky said proved to be key to its performance.
"Because we have the transcriptome for all of the samples, we are able to incorporate that in the training of the prediction model and have an integrated predictor that uses RNA expression levels in a quantitative way to predict the probability of peptide presentation," he said. "It's maybe not surprising in retrospect, but it's that quantitative aspect that [comes from] including the RNA in the training that adds a lot of predictive power over just training on peptides alone."
In a comparison of the EDGE model to a standard binding affinity model, the researchers found EDGE provided up to a ninefold higher positive predictive value.
The team also looked at whether the neoantigens they predicted as likely to be present were, in fact, eliciting an immune response. Using previously generated data in which researchers assessed single-nucleotide variants in 17 patients to determine which had been recognized by T-cells, they found that the EDGE tool identified among its top predicted neoantigens most of the peptides actually being recognized by the T-cells.
Yelensky said that modeling neoantigen binding from the T-cell side might also have value, adding that that was something Gritstone was considering, though it wasn't a part of the Nature Biotechnology study.
The company has patented the model and is using it in a clinical trial of its personalized neoantigen cancer vaccine GRANITE-001, he said. In that work, it collects samples from patients then does tumor exome and tumor transcriptome sequencing to identify the mutations present that might potentially be expressing neoantigens.
"Then EDGE is used to assign a probability of how likely each of those mutations is to be presented on the surface of that tumor in that patient," he said. "We rank those mutations and then the top 20 neoantigens get delivered to the patient as a vaccine."
The company is currently collaborating with Bristol-Myers Squibb investigating the combination of GRANITE-001 with BMS's Opdivo (nivolumab) and Opdivo and Yervoy (ipilimumab) in advanced solid tumors.