I spend most of my days looking at a nanobody sequence and asking one question: once this is inside a body that has no idea it is carrying a therapy, will it find its way to the right cell? That question sounds almost naive until you try to answer it computationally. It is the one that pulled me into in-vivo gene therapy over the last year. It also turned out to be only half the problem, though it took me a while to see the other half clearly.
In-vivo gene therapy is exactly what it sounds like. Instead of pulling cells out of a patient, engineering them in a lab and infusing them back, the therapy is delivered straight into the body and the modification happens where the cells already live. Simpler idea on paper than in practice. For it to work, the delivery vehicle needs a way to pick out its target cell type from billions of others it will pass on the way, and that is the job we have handed to nanobodies. Displayed on the surface of a viral vector or a lipid particle, a nanobody works like a zip code, sending the therapy toward the cell it is meant to reach and mostly leaving everything else alone.
"The nanobody sitting on a vector’s surface is doing more work than it looks like it’s doing. It’s the difference between a therapy that goes everywhere and one that goes exactly where it’s supposed to."
This is why AI-designed nanobodies have become so central to the vectors we are building for in-vivo gene therapy targeting T cells. T cells are the obvious target for a lot of these therapies, but finding them cleanly is its own problem. A few surface markers do most of the work here. CD3 is present on essentially every T cell as part of the receptor complex. CD5 is shared with some B-cell subsets but is still a workable T-cell handle. CD7 is not just a pan-T-cell marker but is also carried by T-cell malignancies like T-ALL, which makes it useful for more than delivery. Each one comes with its own trade-offs in specificity, expression level, and how fast it internalizes once bound.
The catch is that good binders against these targets are not easy to find off the shelf. Nanobodies raised the traditional way, through animal immunization, take months to come back, and even then a lot of them are not tuned for what a delivery vector actually needs from them. De novo design sidesteps that wait. We can now design nanobodies against CD3, CD5 and CD7 from scratch computationally, and the affinity numbers have been the real surprise: several of these de novo candidates bind more tightly than the handful of nanobodies that already existed for these targets. A shortage that used to slow everything down is starting to look like a design choice instead.
CelabGen is a de novo antibody generation platform, which integrates multiple deep learning models to design complete antibody structures and sequences against any target antigen — entirely from scratch, without relying on an existing antibody repertoire. It generates candidate sequences, predicts their three-dimensional structures, and evaluates them against the target antigen, all computationally, so researchers can go from just a target to viable candidates without a wet-lab discovery campaign. This compresses antibody discovery timelines dramatically..
"Targeting isn’t the hardest part of in-vivo gene therapy anymore. Watching a de novo binder outperform something we would otherwise have had to scavenge from a thin public dataset feels like real progress."
But designing a binder that reaches the right cell only answers the first question. The second one showed up the moment we started taking these designs seriously: will the patient’s immune system leave it alone? A de novo nanobody is, by definition, a sequence that has never existed in a human body before. That is the whole point of it, and it is also the risk. If the immune system reads that sequence as foreign, the patient develops anti-drug antibodies, and the therapy loses potency, sometimes quietly over repeat dosing, sometimes all at once. I have watched this concern come up late in enough conversations to believe it belongs much earlier.
This is the work I have been spending most of my time on now. CIMMEXA is Cellogen’s computational immunogenicity prediction platform, and it exists to answer that second question before a candidate gets anywhere near a clinic. It evaluates whether a therapeutic protein is likely to trigger an unwanted immune response by looking at how its peptides are presented across the human HLA landscape, which is not one landscape at all but thousands of variants spread across different populations.
The architectural decision that matters most in CIMMEXA is that it does not trust any single predictor. We integrated three different algorithms and built a majority-voting consensus across more than 11,000 HLA alleles. Individually, these tools are noisy in a specific and annoying way: they flag a lot of peptides that turn out to be nothing. Made to agree with each other, they get much quieter, and what survives the vote is far more likely to be a real liability worth redesigning around. The consensus also brought analysis time down from about five hours to under one, which sounds like a convenience until you are the person waiting on it to decide whether a design round moves forward.
"One algorithm will tell you a peptide might be a problem. Three of them agreeing tells you whether it actually is."
Put next to each other, CellNFo and CIMMEXA are really two halves of the same design loop. One asks whether a molecule can find its target. The other asks whether it will be tolerated once it gets there. Running them in sequence means a candidate can be killed on a workstation instead of in a clinical trial, and the cost difference between those two places to fail is measured in crores and in years.
I do not think targeting is the hardest part of in-vivo gene therapy anymore, not the way it used to be. What is still hard is everything downstream of it: how the payload behaves once it is inside the cell, how the vector holds up through manufacturing, how it behaves in an actual patient instead of a dish. Immunogenicity prediction narrows the uncertainty but does not remove it, and I try to stay honest about that. Still, watching a de novo binder outperform something we would otherwise have had to scavenge from a thin public dataset, and then watching a consensus run tell us which of those designs to be careful with, feels like real progress. It is the part of this work I keep coming back to.
- Research Associate
