The decision a model should support in an LNP programme is narrow: of the candidate ionisable lipids on the list, which ones are worth material and instrument time. That decision does not need an accurate prediction of encapsulation efficiency. It needs a ranking whose differences you can tell apart from noise, a rule for the group at the top that is effectively tied, and the discipline to keep the measurement as the arbiter. This article is about turning predicted numbers into that shortlist.
Decide what counts as a real difference
Start from the model's error on data it has not seen. Suppose the held-out error on encapsulation efficiency is 2.1 percentage points. Two predictions each carry that uncertainty, so the difference between them carries roughly 2.1 multiplied by the square root of two, about 3.0 percentage points. A predicted gap of 1.2 points between two candidates is therefore not a ranking at all, and a gap of 6 points is.
Write that threshold down before you look at the list. It converts an ordered table into three or four groups, which is what the data actually support, and it stops the conversation about whether candidate four or candidate five is better.
Expect the top group to be tied
With 24 candidates and a 3-point resolution, the top of the list is typically five to eight candidates you cannot separate. That is not a failure of the model; it is the honest state of knowledge, and it is useful, because the tie is broken on criteria the model does not cover:
- Synthetic effort. Steps, availability of the amine and tail building blocks, whether the route is one your lab has run.
- Material on hand. A tied candidate you already have in grams beats one that needs two weeks.
- Structural diversity within the tie. Pick candidates that differ in the feature you are trying to learn about, such as the linker or the tail branching, so the run teaches you something beyond which one won.
- Known liabilities. Ester hydrolysis risk, impurity profile, anything the analytical team has flagged before.
A shortlist of four to six, chosen from the tied top group on these grounds, is a defensible plan. A shortlist of the top four by predicted value alone usually tests four versions of the same idea.
The multi-response problem
Encapsulation efficiency alone is the wrong objective. A formulation has to satisfy size, polydispersity and expression as well, and the candidate with the best predicted encapsulation is frequently not acceptable on size. Handle this the same way as any multi-objective problem: state which responses are hard constraints with a window, and which are to be maximised. "Particle size between 70 and 110 nm" and "polydispersity below 0.2" are constraints; expression is the objective. A candidate that violates a constraint leaves the list regardless of its score on the objective.
Lipid identity is not the whole design space
A predicted ranking of lipids is only one axis. The molar ratios of ionisable lipid, helper phospholipid, cholesterol and PEG lipid, the N/P ratio and the process parameters move the responses as much as the lipid choice does, and they are far cheaper to vary. Two consequences:
- Comparing lipids at one fixed ratio set compares them at a point that may suit one and not another, so a lipid can lose on a ratio rather than on its chemistry.
- A tie at the top is often best resolved by taking two tied lipids into a small ratio design rather than by picking one, because the design answers two questions at once.
Close the loop on the first measurements
The first bench results from the shortlist are the most valuable data the programme will generate, for two reasons. They tell you the offset between predicted and measured for your process, which is usually a systematic shift that can be subtracted from the remaining predictions. And they tell you whether the ranking held at all, which decides how much weight the rest of the list deserves.
A worked sequence: shortlist six, run them at one reference ratio set, compare measured against predicted. If the rank correlation is good and the offset is consistent, the remaining eighteen predictions are now usable with a known bias. If the measured values are unrelated to the predicted ones, the model carries no information for this chemistry, and the honest conclusion is to screen the next set by design of experiments rather than by prediction.
Limits of the statement
- Predicted LNP responses are prioritisation. They decide which formulations get bench time and in which order, and they do not substitute for the measured encapsulation, size, polydispersity and expression of the formulation you intend to use.
- Models are built on data from particular processes. A prediction from a microfluidic dataset applied to a different mixing method carries an unknown offset, which is exactly why the first measurements are used to calibrate.
- Novel lipid chemistry is where predictions are weakest and where the interest usually lies. Expect domain flags and plan measurements for flagged candidates rather than arguing with the number.
- Nothing in a prediction addresses tolerability, biodistribution or stability on storage. Those are their own studies.
Frequently asked questions
How large does a predicted difference have to be to matter?
Roughly the model's held-out error multiplied by the square root of two, because each prediction carries that error. With a 2.1-point error on encapsulation efficiency, differences below about 3 points are not a ranking and should be treated as a tie.
What do you do when the top candidates are tied?
Break the tie on criteria the model does not cover: synthetic effort, material already in hand, structural diversity across the tied group, and known analytical liabilities. That produces a shortlist that teaches you something whichever candidate wins.
Should the shortlist be chosen on encapsulation efficiency alone?
No. Size and polydispersity are constraints with windows, and expression is usually the real objective. A candidate outside the size window leaves the list no matter how well it scores on encapsulation.
How many candidates belong in the first bench run?
Enough to measure the model's offset and rank correlation, which in practice is four to six chosen from the tied top group, run at one reference ratio set so the comparison is clean.
What if measured and predicted values turn out to be unrelated?
Then the model carries no information for this chemistry or this process, and the next set should be screened with a designed experiment instead. Finding that out on six formulations rather than thirty is the point of running the comparison early.