I’ve been thinking about longevity since at least 2017. Back then I wrote about it as an investor thesis, the idea that extending human healthspan would become one of the defining technological challenges of our generation. It was somewhat abstract at the time, interesting to think about over dinner and easy to file under “future trends.”
Then I lost two friends to cancer. That changed a lot for me, because something abstract suddenly became personal very fast.
When someone close to you dies from cancer, you go through the usual stages. But at some point, if you’re the kind of person who builds things, you start asking different questions. Not “why did this happen” but “what is actually being done” and “where are the gaps.” I started reading a lot about cancer. First books, then online articles, and eventually I ended up with papers about cancer immunotherapy, personalized medicine, and neoantigen vaccines. I talked to people in the space and then I found a gap that shouldn’t be there.
The cancer vaccine opportunity
What did I learn? Your immune system is designed to find and kill cancer cells. Cancer cells have mutations, and some of those mutations produce tiny protein fragments (called neoantigens) that get displayed on the cell surface like a flag. In theory, your immune system should see those flags and attack them. In practice, however, it often misses them.
A personalized cancer vaccine works by sequencing a patient’s tumor, finding the mutations, predicting which ones will produce flags the immune system can actually recognize, and putting the best ones into a vaccine. The vaccine then trains the immune system to look for those specific flags. BioNTech and Moderna are both running clinical trials on this right now, and some of their early results have been remarkable. For example, patients with pancreatic cancer, one of the deadliest cancers, are now showing durable immune responses.
But there’s a bottleneck. A patient’s tumor might have hundreds of mutations, but only a handful will produce neoantigens that actually trigger an immune response, while the vaccine can only carry about 10 to 20 candidates. So the question is: which ones do you pick? I realised this is a computational problem and that the current tools are not solving it well.
The gap in the market
The standard tool used in clinical neoantigen prediction is called pVACtools. It was developed at Washington University and published in 2020. It’s a good tool, it’s open source, well-maintained, and widely used. But it ranks neoantigen candidates primarily by just one thing, how strongly the peptide binds to the patient’s immune presentation molecules.
But of course it’s not the whole picture, because there are many biological factors that pVACtools doesn’t account for. Is the gene actually turned on in the patient’s tumor? If not, the mutation can’t produce a neoantigen, no matter how strong the binding. Does the mutation make the peptide look sufficiently “foreign” to the immune system compared to the normal version? Is the mutation in a position where T-cells can actually see it?
Published research has shown these factors matter. The TESLA consortium (a group of 36 research teams that has nothing to do with Tesla, TESLA stand for The Tumor Neoantigen Selection Alliance) identified five key parameters governing neoantigen immunogenicity. Multiple academic groups have demonstrated that machine learning models incorporating these features outperform binding-only ranking. Yet in clinical practice, most vaccine trials still use binding affinity as the primary or only ranking signal. I saw an opportunity to build something better.
Sean Clark, Helixion Therapeutics and Elyra
Sean Clark has been my business partner on all my ventures for the past 10 years. Together we started a number of successful businesses in the past, among others Nasdaq-listed Hut 8, a Bitcoin Trust that turned into a TSX-listed ETF, First Coin Capital that was sold to Galaxy Digital as part of its public listing process, and a number of private companies and funds. We now founded Helixion Therapeutics together in Vancouver.
Sean is a builder, someone who just gets things done. He is not a career academic, but someone with a deep technical ability who can take a complex problem and ship a working solution. We complement each other well, I bring the strategic vision and network, he brings the engineering execution and is one of the best sales people I ever worked with.
We built Elyra, a neoantigen prediction engine that takes tumor sequencing data and produces a ranked list of vaccine candidates. The scoring model uses five biologically meaningful features: gene expression, binding affinity, agretopicity (how much stronger the mutant peptide binds compared to the normal version), peptide length, and the position of the mutation within the peptide.
The most important finding during development was that gene expression is the single strongest predictor of neoantigen immunogenicity, even more important than binding affinity! It accounts for 36.3% of the model’s predictive power, compared to 29.5% for binding. This makes biological sense: if the gene isn’t active in the tumor, nothing downstream matters. But pVACtools doesn’t use expression for ranking. And that’s the gap we exploit.
We also learned, the hard way, that training data matters more than training volume. An early version of our model that was trained on 47,000 peptide samples (including infectious disease data) actually performed worse than a model trained on 18,000 cancer-specific samples. The infectious disease data was diluting a cancer-specific signal called agretopicity that turned out to be one of our most important features. Sometimes less data is better, if it’s the right data.
External validation on the TESLA benchmark
In this field, everyone claims their model works. The only way to know for sure is to test on truly external data, meaning testing it on patient data that the model has never seen during training.
The TESLA benchmark is the field’s gold standard for this. It was created by 36 research groups and comprises 918 peptides across 9 patients with melanoma and non-small cell lung cancer, with experimentally confirmed immunogenicity labels. Basically, if your model works on TESLA, it works in the real world.
Elyra achieves a Global AUC of 0.812 on TESLA. That beats pVACtools (0.797) and matches the TESLA consortium’s own published composite model. Precision@10, which measures how many of the top 10 candidates are genuinely immunogenic, improved more than three times compared to binding-only ranking.
To be transparent: the margin over pVACtools is real but not enormous. We’re preparing a preprint with a full statistical analysis. But the direction is clear, and the model has room to improve. We haven’t yet integrated three-dimensional structural prediction (we have Protenix-v1, an open-source model that matches AlphaFold3, ready to go), and we’re working on incorporating real patient-level RNA-seq expression data to replace the population-level proxy we currently use.
Why now?
I believe the timing is right for three reasons.
First, the underlying science has matured. Cancer neoantigen vaccines have gone from theoretical to clinical. BioNTech, Moderna, and others are running Phase II trials. The results so far are encouraging enough that more trials are coming. Every one of those trials needs neoantigen prediction, and the current tools haven’t kept up with what we now know about immunogenicity.
Second, the open-source ecosystem has reached a tipping point. Structure prediction (Protenix-v1), binding prediction (MHCflurry), epitope databases (CEDAR, IEDB), the components needed to build a serious neoantigen pipeline, are now all publicly available under permissive licenses. A few years ago, building this would have required either a large academic lab or a well-funded biotech with proprietary tools. That barrier is almost completely gone.
Third, and this is the one most people in biotech have not fully absorbed yet: AI has changed what a small team can build. AI allows you to dramatically accelerate the implementation of ideas. The bottleneck used to be that even if you knew what to build, the engineering took months or years. That’s no longer the case. The bottleneck now is knowing what to build, things like domain insight, judgments about which features matter, which datasets to trust, and which validation to run. That’s the hardest part, and that’s what Sean and I bring to the table.
The vision
Elyra’s scoring model is one component of a larger pipeline. The full vision is an end-to-end service: a research partner or biotech sends us tumor sequencing data, and we return a validated, ranked neoantigen report with structural analysis and expression profiling. Think of it as neoantigen prediction as a managed service, the computational biology department that small vaccine companies don’t have and can’t afford to build.
We plan to integrate Protenix-v1 for surface exposure analysis (determining whether neoantigens are physically accessible to the immune system), real RNA-seq expression data, and eventually additional validation across more cancer types. A preprint is in preparation and will be posted on bioRxiv.
We’re looking for a scientific advisor
This is where I need help from my network. We are looking for one or more computational biologists or immunologists to join Helixion as a scientific advisor.
The right person knows the neoantigen prediction space from the inside. They’ve worked with tools like pVACtools, NetMHCpan, or MHCflurry. Ideally they have hands-on experience with T-cell assays, cancer vaccine clinical trial design, or immunogenomics pipelines. They may have published in this area.
But just as importantly, they’re entrepreneurial. This is a small company at an early stage. The company is officially based in Vancouver, Canada, but we work from all over the world (I am mostly based in Singapore for example). We are pre-revenue and we are building something from near-zero. If that excites you, we should talk. If you need a big team and a guaranteed salary, this probably isn’t the right fit.
Starting as a advisory role with stock options in the company at a very low strike price. If there is chemistry, we can discuss something bigger.
If you know someone who might be interested, I’d appreciate an introduction. And if you’re that person yourself, reach out to marc@helixiontx.com or DM me on X (@marcvanderchijs).
A personal note
I started this post talking about losing friends. I want to end there too. The reason I care about this is not because it’s a good investment thesis, although I think it is. It’s because cancer is personal for almost everyone. If Elyra can help put even slightly better neoantigens into a vaccine, that translates directly into a stronger immune response in a real patient. Maybe that patient gets a few more years. Maybe the cancer doesn’t come back.
I don’t know if Helixion will become a large company. I don’t know if Elyra will become the standard tool in this space. But I know the problem is real, the science is ready, and the tools exist to build something meaningful. I’d rather try and fail than watch from the sidelines while the technology catches up to what we already know.
If you’ve read this far and any of it resonates, get in touch.
Thanks for sharing. Quite interesting. I have also learned a lot about the disease and immuno..