Empirical screening, not ML, designs SV-Delivery™ kits
How SunVax explicitly chose an empirical-then-pattern workflow for LNP discovery, and why that matters for buyers reading "AI-driven" copy elsewhere.
When LNP vendors describe their design pipeline, you'll hear two stories. One story says "we use machine learning and computational design to predict the best ionizable lipid." The other says "we made thousands of variations, screened them on real cells, and noticed which combinations worked." SunVax is firmly in the second camp — and they say so on the record.
At the Antibody 2026 panel on 2026-05-15, Yingzhong Li (SunVax founder & president) was asked whether their LNP discovery process leans on machine learning. His answer was direct: "we hardly find any useful how to design that. And then we use the very general design methods and then we find some pattern." When the follow-up question pressed — "so you start from pure empiricism and later on you build your own database?" — his reply was: "Yeah, yeah, yeah. Yes, so that's correct."
This isn't a hedge. It's a methodology decision. The LNP chemical space is roughly 10¹⁰ candidates if you only vary the headgroup, linker, tail length, and PEG content. Public LNP datasets have on the order of 10³–10⁴ data points. Training a model on 10³ data and asking it to extrapolate to 10¹⁰ is bad statistics, not good engineering.
What SunVax does instead, as Li described in five steps on the same panel: (1) design a small formula library tuned to a target cell type; (2) test it; (3) pick the top group and design the next round; (4) move best candidates into target-cell assays with activation-state controls; (5) run mouse biodistribution and pick a winner. Multiple rounds. Thousands of data points accumulated over time. Patterns emerge from accumulation — not from first-principles prediction.
For a buyer, this matters. If you're spending $5,000 on an SV-Delivery™ kit and another vendor tells you they "AI-designed" their LNP from theory, you should ask: what was the training set? what was the validation set? what data were withheld? If the answer is hand-wavy, the kit you receive is probably empirical too — they just don't say so.
We do not market sunvax.x1000.ai as "AI-driven LNP design", because SunVax doesn't claim that, and we will not put a phrase on this site that the founder has explicitly disclaimed. This is what evidence-anchored marketing looks like in practice.