Parker Institute summoned the neoantigen pioneers to compare their prediction algorithms. And they have some suggestions for the field
What makes a good neoantigen?
For all the promises of the bold new approach to cancer vaccines and therapies — zeroing in on specific mutated antigens expressed only by tumors — companies and academics have little way of knowing how good they are at predicting which neoantigens represent the best targets. There’s no standard or baseline for players to stack themselves against rivals in the nascent field, and by the time they find out, it could be too late.
No single group could really build that benchmark. Precious proprietary information is at stake, not to mention tremendous resources required.

But four years ago, just as the concept was taking off, it struck the Parker Institute for Cancer Immunotherapy as the exact kind of problem its collaborative model was built to solve.
“We’re like Switzerland of neoantigens,” Danny Wells, the principal data scientist at PICI, told Endpoints News.
So working with the Cancer Research Institute and a nonprofit named Sage Bionetworks, it brought together over 40 biopharma companies and academic labs, gave them the same melanoma and non-small cell lung cancer tissue, and asked each team to submit its most promising neoantigen predictions. PICI researchers then went into the lab and cross-compared the predictions, checking whether the neoantigens were indeed recognized by T cells. The result is a baseline dataset that the initiative — named TESLA, short for Tumor Neoantigen Selection Alliance — is making public to the scientific world today.
To the researchers’ surprise, the differences between the prediction algorithms were “tremendous,” said Wells, a co-senior author in the paper published in Cell. No team managed to identify every neoantigen or even a large majority of them: “The overlap between predictions, no matter how we sliced it, was really low.”
It highlights the need for new knowledge in the field, he added, and PICI believes TESLA has yielded some insights.

A set of five distinct features, it turned out, could predict good neoantigens with high accuracy and specificity when integrated into a model. They are binding affinity, tumor abundance and binding stability, which has to do with how the neoantigens are presented; as well as agretopicity and foreignness, which relates to their recognition by immune cells.
“These are all features that had been talked about, but I think we were surprised that just by integrating them together into this single model that it works so well,” Wells said.
When participating teams reapplied these characteristics into their algorithms, PICI reported, the predictions improved. A data model emphasizing all five features came out of a test against another set of cancer samples accurately predicting 75% of effective neoantigen targets and filtering out 98% of ineffective ones. Depending on the therapeutic strategy drug developers may calibrate their algorithms differently, Wells said. But the hope he and co-senior author Nadine Defranoux have is that they can provide a common baseline both for those already in the field and others looking to jump in.

“This research has the potential to improve drug makers’ and researchers’ mathematical algorithms,” Lisa Butterfield, vice president of research and development at PICI, said in a statement. “It can prioritize antigens most likely to be present on each patient’s cancer and most visible to the immune system while deprioritizing the ones that aren’t.”

While TESLA won’t be monitoring the space longitudinally to see how companies compare over time, Wells does envision new evolutions for the coalition. Many questions still need to be answered; for instance, while they focused on class 1 prediction, or how CD8+ T cells see the tumor, co-senior author Robert Schreiber’s research has suggested that class 2, or how CD4+ T cells see the tumor, are just as important.
“Until now, neoantigen prediction has been a black box,” said Schreiber, a professor at the Washington University School of Medicine in St. Louis. TESLA has begun shedding light on it, and it intends to continue.