A model trained on published LNP data usually ranks your formulations better than it predicts them. The ordering survives the move to your lab; the absolute values do not, because your mixing geometry, flow rates, buffer, mRNA construct and assay all shift the numbers. Fine-tuning on your own measurements is what removes that shift, and it needs far less data than building a model from scratch. The work is mostly in the dataset and the split, not in the training.
Why the public model is offset on your process
Encapsulation efficiency, particle size and polydispersity are process responses as much as formulation responses. Three sources of offset, in rough order of size:
- Mixing. Microfluidic, T-junction and pipette mixing give different particle populations from the same lipid composition, and the difference is not a constant for all lipids.
- Assay. A dye-exclusion readout calibrated differently, or a different standard curve, moves encapsulation efficiency by a few percentage points systematically.
- Payload. Construct length and sequence change the N/P ratio that works and therefore the response surface.
The useful consequence is that the offset is largely systematic, so a modest amount of in-house data can correct it, and the ranking information from the public data is retained.
What a usable in-house dataset contains
Per formulation, the following must be recorded in machine-readable form, and the third and fourth are the ones most often missing:
| Field | Note |
|---|---|
| Ionisable lipid identity as a structure | A code name alone cannot be modelled |
| Molar fractions of all four lipid components, and N/P | Fractions that do not sum consistently are a frequent data error |
| Process parameters: mixing method, flow rates, buffer, pH, temperature | Without these, runs from different weeks are not comparable |
| Assay details: method, standard, limit of quantitation | Needed to know whether a value is a measurement or a bound |
| All responses measured, including the failures | A dataset of acceptable batches cannot learn where the edge is |
| Batch or run identifier and date | The basis for an honest split |
Fifty to two hundred formulations with these fields complete are worth more than five hundred rows with the process parameters missing.
Split by batch and by lipid, never at random
A random split of formulation data is the standard way to get an encouraging and meaningless validation figure, because a near-identical formulation from the same run lands on both sides. Two splits that answer real questions:
- By run or batch. Hold out whole runs. This measures whether the model transfers across day-to-day process variation, which is the question when you want to use it next week.
- By ionisable lipid. Hold out all formulations containing certain lipids. This measures whether the model generalises to a lipid it has not seen, which is the question when you want to prioritise new candidates.
Expect the second to look worse than the first. That is not a defect in the model; it is the difference between interpolating ratios and extrapolating chemistry, and both numbers belong in the report.
Start with calibration, not with training
Before fine-tuning, check whether a two-parameter correction is enough. Plot measured against predicted for twenty of your formulations and fit a straight line. If the relationship is tight with a slope near one and a constant offset, you do not need a new model: you need to subtract the offset. This takes an afternoon and sometimes ends the project satisfactorily.
Fine-tuning earns its place when the residuals are structured, for example when one lipid class is systematically over-predicted and another is not. That is a pattern a constant cannot fix.
Judge the result on decisions, not on error metrics
A fine-tuned model with a slightly better error can still be useless, and a modest one can be valuable. Three checks that correspond to how the model will be used:
1. Rank correlation on a held-out run. Does it order the formulations in that run correctly? This is what a shortlist depends on. 2. Top-k recall. Of the five best formulations in the held-out set, how many are in the model's top five? A model that finds three is doing the job. 3. Constraint classification. For responses with a window, such as particle size, how often does it put a formulation on the correct side of the limit? Being right about the window matters more than being right about the value.
Report all three with the held-out set described. A single error figure without the split rule says nothing, and a figure from a random split overstates performance in a way that becomes visible only at the bench.
Keep the model current
Processes drift, methods get revised, new lipid classes enter the programme. Two habits keep a fine-tuned model honest: add each new run to the dataset with its process parameters, and re-measure the calibration quarterly on the most recent runs. A model that was good a year ago and has not been checked since is an unexamined assumption in every prioritisation it has informed.
Limits of the statement
- Fine-tuning corrects for your process and your assay. It does not make the model valid for a process you have not measured, such as a scale-up to a different mixer.
- Predicted responses remain prioritisation. The formulation you intend to use needs its encapsulation, size, polydispersity and expression measured.
- A model fine-tuned on one payload class should be treated as unverified for a different one until measurements say otherwise.
- None of this addresses in vivo behaviour, tolerability or stability on storage. Those are separate studies with their own designs.
Frequently asked questions
Why does a published LNP model not match our numbers?
Because encapsulation efficiency, size and polydispersity depend on the process as well as the formulation. Mixing method, assay calibration and payload each shift the values, mostly systematically, which is why a modest in-house dataset can correct the offset.
How many of our own formulations are needed?
Fifty to two hundred with complete process parameters is a typical useful range for fine-tuning, and twenty are often enough just to measure and remove a systematic offset. Completeness of the fields matters more than the row count.
How should formulation data be split for validation?
By run and by ionisable lipid, not at random. A random split puts near-identical formulations from the same run on both sides and produces an encouraging figure that does not survive the bench.
Is fine-tuning always necessary?
No. If measured against predicted is a tight line with a slope near one, subtracting the offset is enough. Fine-tuning is for structured residuals, such as one lipid class being consistently over-predicted.
What metric should decide whether a fine-tuned model is good?
The ones that match its use: rank correlation on a held-out run, how many of the true top five appear in the predicted top five, and how reliably it places formulations on the correct side of a constraint window.