The problem

Lipid nanoparticles (LNPs) are the delivery vehicle behind mRNA therapeutics — including the COVID-19 vaccines. Every new formulation is characterized by its encapsulation efficiency (EE%): the fraction of input nucleic acid that ends up inside the particle rather than free in solution. High EE is one of the first go/no-go metrics in LNP development, and it is measured empirically for nearly every new formulation.

That empirical bottleneck is the target here. If a predictive model can rank-order candidate formulations before synthesis, wet-lab teams can front-load the promising ones and skip the obvious losers. Even an imperfect model that narrows the search space is useful.

The catch is data. Published LNP literature is heterogeneous: different labs report different subsets of formulation parameters, assay conditions vary, and EE values reflect different measurement protocols. The signal is there, but so is a lot of noise.

Approach

I'm developing LNP-EE-Predictor as an open-source pipeline in collaboration with Constantinos Skevofilax, trained on the LNP Atlas dataset — 1,092 formulations curated from peer-reviewed literature. The project explores two complementary strategies:

  • XGBoost baseline — a high-capacity gradient-boosted tree that extracts signal from rich feature engineering across all 636 usable rows.
  • Physics-informed neural network (PINN) — a smaller network trained on a 93-row subset with the full feature vector, plus three physics residual terms encoding domain constraints.

Both models are evaluated the same way: GroupKFold cross-validation grouped by paper DOI. Random-split validation on this data is misleading because rows from the same paper share confounders — measurement protocol, buffer composition, operator technique, lipid lot. GroupKFold forces the model to generalize across papers, which is what matters if you want to use it on a new formulation from a new lab.

Results

Model OOF R² OOF RMSE Rows Features CV strategy
XGBoost −0.032 21.88% 636 88 GroupKFold (paper DOI)
PINN, 7-feature 0.438 11.25% 93 7 GroupKFold (paper DOI)
PINN, 7-feature 0.580 10.12% 93 7 Random 80/20 (leakage-prone)

The honest number is the middle row: R² ≈ 0.44 on out-of-fold cross-paper prediction using the PINN. The XGBoost baseline is worse than the mean predictor on cross-paper splits, which is a real and instructive result — not a bug to hide.

Why is XGBoost R² negative?

A previous iteration reported XGBoost R² ≈ 0.136. That number included a data leakage bug: the cleaned target column (encapsulation_efficiency_clean) was inadvertently retained as an input feature after the EE-cleaning step. Fixing the leak dropped honest R² to −0.032.

Both the XGBoost and the PINN results reveal the same underlying difficulty: cross-paper generalization from heterogeneous LNP literature is hard. Root-cause analysis suggests:

  • Within-paper EE variance ≈ between-paper EE variance (both ~0.02 in fraction units)
  • Roughly half of the EE variance is explained by unmeasured, paper-level confounders that no formulation feature can capture
  • Feature coverage is uneven — many published papers omit key formulation details, so the model has to impute or drop

Dataset cleaning

Getting the dataset from 523 usable rows to 636 involved:

  • Fixing 174 rows with corrupt ± encoding in the source CSVs
  • Parsing range-formatted EE values (e.g. "80–85%") and dual-assay reports
  • Correcting 6 broken cKK-E12 SMILES against PubChem CID 71555845
  • Nulling 32 DSPE-2armPEG2000 SMILES that had valence errors (falls back to frequency encoding)

Cleaned data is committed as data/lnp_atlas_cleaned.csv — the goal is that anyone reproducing the work starts from the same numbers.

Physics-informed neural network

The PINN adds three physics residual terms (R1, R2, R3) to the loss function, encoding first-principles constraints on how formulation parameters should relate to EE. On the 93-row subset with full feature coverage, this halved RMSE relative to a comparable MLP without physics constraints. The tradeoff is the small labeled dataset — the physics prior is doing real work compensating for data scarcity.

Expanding the PINN training set is the next major work item. Every paper that reports a full formulation vector unlocks new training examples.

Interactive dashboard

The Streamlit dashboard exposes data exploration, both model comparisons, and an interactive predictor where you can enter a formulation and get a predicted EE% back:

Hosted on Streamlit Community Cloud, which sleeps after inactivity — the first load can take 30–60 seconds. Open the dashboard in a new tab →

What's next

  • API endpoint — Dockerized FastAPI service so the model is programmatically callable, not just interactive.
  • Expand the PINN dataset — the 93-row subset is the current ceiling; growing it toward 300–500 rows would likely push honest R² past 0.6.
  • Public predictions log — every prediction requested through the API gets logged (formulation + predicted EE), so if a user later measures the actual EE and reports it, we can grow the dataset from real usage.