Predicting where a lipid nanoparticle goes, before you make it

Lipid nanoparticle development has traditionally run on a make-dose-measure loop: formulate a composition, inject it, image where it ended up, then change one thing and repeat. Within a defined applicability domain, that loop can now be replaced with a model. Change the ionisable lipid content, the helper phospholipid, the cholesterol, the PEG-lipid, or the particle size, and watch the predicted in-vivo delivery move, across eight tissues, in real time, in a browser.

Why fourteen formulations, not one at a time

Fourteen lipid nanoparticle compositions were chosen by Design of Experiments, not by intuition or by testing one variable while holding the others fixed. A one-factor-at-a-time approach can't see interactions: cases where the effect of PEG-lipid content depends on which ionisable lipid it's paired with, for instance. A designed space covers the composition space efficiently enough that those interactions become visible, from a number of formulations small enough to actually dose.

Each of the fourteen formulations was imaged in vivo, in eight tissues: liver, spleen, kidney, bone marrow, lung, heart, brain and whole blood. That's the entire experimental dataset the model was built from. Everything the profiler shows you for a composition that wasn't one of the fourteen is a prediction, interpolated by a self-validated ensemble model trained on that designed set, not a new animal study.

Reading the profiler

Each small panel is one input: ionisable lipid content, DSPC, cholesterol, DMG-PEG, particle size. The trace inside it is the model's prediction as that single input varies, holding everything else where you've set it. Drag the vertical line and every trace across every tissue redraws at once, because the inputs interact with each other, not just with the outcome.

• A steep trace means the prediction is sensitive to that input in that tissue.
• A flat trace means it barely matters there, which is often the more useful finding.

Why interactive, not just published

A model that lives only in a figure can't be interrogated. It shows the formulations the authors chose to plot, not the one a reader is actually trying to design. Making it interactive means someone with a real formulation question can test it directly, and it does the thing the underlying paper argues for: answering a question computationally, in seconds, that would otherwise have needed another in-vivo study.