Decoding Extrahepatic Targeting of Lipid Nanoparticles with Interpretable Machine Learning
Asal Mehradfar, Mohammad Shahab Sepehri, Owen Antholine, Varun Shankar, Glen S. Kwon, Salman Avestimehr, Morteza Rasoulianboroujeni
q-bio.QM · cs.CE · cs.LG
Quick verdict
ML on 476 curated LNPs predicts liver vs extrahepatic IVIS outcomes (AUC≤0.874) and gives design rules; study-level split leakage may inflate results.
Summary
The authors curate a literature-derived dataset of 476 intravenously administered, ligand-free LNP formulations from 81 studies (2011-2025), labeled hepatic vs extrahepatic by the organ with maximal IVIS signal (239 liver / 237 non-liver). Each formulation is encoded as an 808-dimensional vector: RDKit Expert descriptors computed from standardized SMILES of six component classes (ionizable lipid, helper lipid, sterol, PEGylated/polymer lipid, additional lipid, polymer repeat unit) plus mole fractions and polymer molecular weight. Logistic regression, random forest, and XGBoost reach held-out test ROC-AUCs of 0.839/0.866/0.874 (5-fold CV 0.826-0.882). SHAP analysis with a median-rank consensus across models identifies ionizable-lipid descriptors (VSA_EState5/7, MinEStateIndex, BalabanJ) and composition variables (IL%, SL%, PL%) as dominant predictors; retraining on the top 20 consensus features recovers random-forest performance (0.866). The authors derive design rules (higher IL%, lower PL%, amide motifs, unbranched alkyl segments, high IL-VSA_EState7) and qualitatively corroborate the compositional trends with an excluded study (Zhu et al.). Main limitations: the random formulation-level split risks study-level leakage (one paper contributes 65 formulations), binary max-signal labels ignore dose/imaging-time heterogeneity and pool heterogeneous organs, no composition-only baseline isolates the descriptors' incremental value, and no statistical comparison supports the model ranking.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 4
● low 0
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An 808-dimensional representation of lipid SMILES descriptors plus formulation composition predicts hepatic vs extrahepatic accumulation of IV-administered LNPs with held-out test ROC-AUC up to 0.874 (XGBoost).
- Ionizable-lipid molecular descriptors dominate prediction, with formulation composition (ionizable lipid, sterol, and PEGylated-lipid mole fractions) contributing substantially.
- The top 20 consensus-ranked features retain nearly all predictive information, with random forest matching its full-feature test AUC of 0.866.
- Extrahepatic accumulation is associated with higher ionizable-lipid fraction, lower PEG-lipid fraction, higher IL-VSA_EState7, and amide and unbranched-alkane motifs, yielding descriptor-guided design rules for non-hepatic LNPs.
Strengths
Transparent, criteria-based dataset curation
Section 2.1 states four explicit inclusion criteria (IV administration, multi-organ IVIS data, no active targeting ligands, fully extractable structures), and Table S1 lists all 81 source DOIs with per-paper formulation counts, enabling audit of the 476-formulation dataset.
Interpretability claims are empirically validated, not just asserted
Rather than only reporting SHAP rankings, the authors retrain all three models on the top 20/10/5 consensus features (Table 1, Table S4, Figure S2); random forest matches its full-feature test AUC (0.866) with only 20 features, directly demonstrating that the consensus set carries most predictive information.
Cross-model consensus design reduces single-model artifacts
Calibration and external consistency checks beyond AUC
Weaknesses
Study-level data leakage risk in the evaluation split
Weak and heterogeneous label definition
Missing composition-only baseline
No statistical inference on model differences
Speculative mechanistic reading of coarse binned descriptors
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How accurate is this machine learning model for predicting LNP biodistribution?
On a held-out test set of roughly 95 literature-derived formulations, the models classify LNPs as liver- vs non-liver-accumulating with ROC-AUC of 0.839 (logistic regression), 0.866 (random forest), and 0.874 (XGBoost), with accuracies around 78-85%. Because the train/test split is random at the formulation level and many formulations come from the same papers, independent experts may consider these figures somewhat optimistic.
What design rules does the paper propose for extrahepatic (non-liver) LNP delivery?
Based on SHAP feature directions, the paper recommends a relatively high ionizable-lipid molar fraction, a low PEGylated/polymer-lipid fraction, ionizable lipids containing amide motifs and unbranched alkyl segments, and favorable electrotopological surface profiles (high IL-VSA_EState7 values). These are correlational, descriptor-level guidelines that still require experimental validation and are not organ-specific (lung vs spleen vs other).
Where does the training data come from and how reliable are the labels?
The dataset contains 476 intravenously administered, ligand-free LNP formulations extracted from 81 published studies (2011-2025). A formulation is labeled 'liver' or 'non-liver' depending on which organ showed the highest IVIS imaging signal. This is a literature-derived proxy rather than a standardized measurement, so labels inherit heterogeneity in dose, reporter, imaging time point, and normalization across studies.
Is the code and dataset from this paper publicly available?
Yes. The authors state that both the code and the curated dataset are publicly available at github.com/AsalMehradfar/lnp-extrahepatic-targeting, and the paper reports fixed random seeds and full hyperparameters, which supports reproducibility of the modeling pipeline (though the hyperparameter tuning procedure is not specified).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (ROC curves, SHAP summary plots, calibration curves, dataset overviews) are not analyzable from the extracted text; directional feature claims rely on the paper's own narration of Figure 5 and Figure S2.
- The public GitHub repository was not inspected, so dataset/code availability is taken from the paper's statement.
- This is an arXiv preprint (2609.17721) and has not yet undergone journal peer review.
- Minor sections or supplementary detail may be truncated in extraction, but the main text, Table 1, and supplementary tables S1-S4 appear complete.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.