---
title: "Choosing and confirming follow-up recipes in a 23-run lipid screen"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Choosing and confirming follow-up recipes in a 23-run lipid screen}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

The worked example fits elastic-net SVEM, forward-likelihood SVEM, GauPro and Extra Trees on the same experimental factors. The installed package
includes the complete executable analysis, so the results can be regenerated
with different measurements or a different preparation budget. To rerun it,
copy the source to a writable project directory and render it there. Install
the suggested packages `GauPro` and `ranger` to include both companions.
The final sections use a two-factor teaching surface and an optional general
spline on the full mixture/process experiment with its nominal ingredient
factor. They show explicit model and exploration counts. The worklist reports how many distinct settings those models
nominate. Install the suggested package `mgcv` for that section.

The journal articles are the fixed versions of record. The authors occasionally
update the arXiv manuscripts. These links open the latest posted revisions.

* Karl et al., *A Workflow for Lipid Nanoparticle (LNP) Formulation Optimization*.
  [Latest arXiv version](https://arxiv.org/abs/2212.11264) · [JoVE article](https://doi.org/10.3791/65200).
* Karl, *SVEMnet: An R Package for Self-Validated Elastic-Net Ensembles and Multi-Response Optimization*.
  [Latest arXiv version](https://arxiv.org/abs/2511.20968) · [Chemometrics article](https://doi.org/10.1016/j.chemolab.2026.105660).

```r
source_file <- system.file("examples", "lipid-screen-follow-up.Rmd", package = "SVEMnet")
file.copy(source_file, "my-lipid-follow-up.Rmd")
rmarkdown::render("my-lipid-follow-up.Rmd")
```

The GP fits in this example use one seeded optimizer start per endpoint.
Rerunning with a different seed can lead to a different fitted surface and
recipe nomination. Compare plausible nominees in the laboratory when that
sensitivity affects the decision.

For a large recipe table, the optional medoid step compares pairs of
high-spread recipes. Its memory cost grows with the square of the shortlist
size. A discrete table of settings the lab can prepare keeps this step small.

```{r worked-results, echo=FALSE, results='asis'}
snapshot <- system.file("examples", "lipid-screen-follow-up.md", package = "SVEMnet",
                        mustWork = TRUE)
cat(readLines(snapshot, encoding = "UTF-8"), sep = "\n")
```