Multi-Objective Active Learning Overcomes Liver Tropism in LNP- Mediated mRNA Delivery to the Brain

<title>Abstract</title> <p> Lipid nanoparticles (LNPs) are the leading platform for mRNA delivery, yet their intrinsic liver tropism—driven by endogenous apolipoprotein E (ApoE) adsorption and hepatic LDL-receptor uptake—largely excludes them from the central nervous system (CNS). Functionalizing LNPs with ApoE-mimetic peptides to engage receptor-mediated transcytosis is a biologically grounded strategy, but the combinatorial formulation space is far too large for purely empirical screening. We present a reproducible, CPU-accessible machine-learning framework that (i) learns interpretable transfection surrogates from public LNP data, (ii) generates an information-optimal formulation library to seed experiments, and (iii) navigates the design space with a multi-objective active-learning loop that explicitly penalizes predicted liver uptake. On the public AGILE library (1,200 combinatorial ionizable lipids), gradient-boosted surrogates predicted HeLa transfection with R ²  = 0.69 and Spearman ρ = 0.83 under random cross-validation; a stricter leave-head-group-block-out evaluation revealed that generalization is combination-level rather than block-level (R ²  ≈ 0), an honest boundary on extrapolation. SHAP analysis recovered chemically intelligible design drivers (hydrogen-bond donors, specific tail/head building blocks, molecular weight, lipophilicity). Retrospective active learning recovered substantially more of the maximum library potency than random screening at equal budget (94% versus 84% after 100 evaluations), concentrating effort on the strongest lipids. We further illustrate a multi-objective extension that co-optimizes brain delivery against liver de-targeting on a mechanistically motivated benchmark. The framework runs end-to-end on commodity hardware and is designed to ingest human-relevant blood–brain-barrier (BBB)-on-chip data as they become available. Code and data are openly released. </p>