The study’s central advance was its validation design rather than the introduction of a new model. Nanoparticle biodistribution records often contain repeated measurements from the same formulation at different time points. A conventional row-wise split can therefore place closely related measurements in both training and evaluation sets, making performance appear more generalizable than it is. The researchers instead grouped records by the database’s nanoparticle identifier, NM.ID, so every measurement associated with one nanoparticle remained within a single partition. Their fixed hold-out set comprised 1,872 training records from 385 nanoparticle identities and 447 test records from 97 identities, with no identifier overlap. Hyperparameters, categorical preprocessing and ensemble weights were fitted using training data alone, while the hold-out outcomes were reserved for final evaluation.

On the untouched hold-out set, the weighted ensemble produced the lowest mean absolute error, 1.8320 percentage points, while the stacking ensemble’s RMSE was only 0.0025 lower than the weighted ensemble’s 3.8132. Random Forest was close behind, with an RMSE of 3.8149 and an R² of 0.0881. LightGBM reached an R² of 0.0752, and XGBoost reached 0.0336. The ensembles therefore offered only marginal gains over Random Forest. More demanding five-fold cross-validation grouped by nanoparticle identity gave similarly modest results: XGBoost had the highest mean R² at 0.2280, with a 95% confidence interval from 0.0011 to 0.4549, followed by Random Forest at 0.2169 and LightGBM at 0.1980. The wide, overlapping intervals did not establish a meaningful advantage for any model.

Performance also varied when the researchers repeated the grouped train-and-test procedure 30 times. Across those partitions, Random Forest achieved a mean R² of 0.1419, XGBoost 0.1055 and LightGBM 0.0924, but the model ranking changed with the nanoparticle identities assigned to each subset. The results indicate that a single favorable split can obscure substantial variation in how models handle unfamiliar formulations. An even stricter analysis grouped the data by literature reference, keeping records from the same publication together. In that cross-study setting, all three base models achieved negative mean R² values, meaning that they did not outperform a mean-response baseline on average when evaluated against entirely separated sources. These findings underscore how difficult it is to transport predictions across studies that differ in formulations, tumor models, protocols, pharmacokinetic conditions and measurement procedures.

SHapley Additive exPlanations, or SHAP, were used to inspect the final tuned XGBoost model. After contributions from one-hot encoded categories were aggregated back to their original variables, animal model had the largest mean absolute SHAP value at 0.4762, followed by tumor cell type at 0.4272, core material at 0.3080, administered dose at 0.2955 and zeta potential at 0.2680. Time point, particle size, polydispersity index, shape and tumor size had progressively smaller contributions. These values describe associations learned by this particular fitted model; they do not establish biological mechanisms or causal effects. The small contribution assigned to tumor size therefore describes its role in the model, not its overall biological importance in nanoparticle transport.

Administered dose required particularly careful interpretation because the prediction target was tumor accumulation expressed as percentage of injected dose, or %ID. Dose had essentially no marginal linear or monotonic association with the target in the analyzed data. Yet removing it from the models weakened five-fold nanoparticle-grouped cross-validation performance: mean R² declined from 0.2280 to 0.2108 for XGBoost, from 0.2169 to 0.2007 for Random Forest and from 0.1980 to 0.1236 for LightGBM. The authors interpreted this as modest contextual information supplied through interactions with other variables, not as evidence that dose independently or causally determines tumor %ID. Similarly, the strong SHAP contributions of animal model and tumor cell type identify predictive patterns without establishing why one biological setting differs from another.

Generalization was also weaker for nanoparticle core materials absent from the training data. Of 447 final hold-out records, 53, or 11.86%, belonged to 11 material categories not represented during training, including iridium, platinum, ruthenium, silica-polymeric materials, polymeric liposomes and silver. The weighted ensemble’s mean absolute error rose from 1.7299 for seen materials to 2.5908 for unseen materials, while its RMSE increased from 3.6325 to 4.9543. XGBoost showed the same pattern, with MAE increasing from 1.7498 to 2.6896 and RMSE from 3.7093 to 5.2712. Although LightGBM produced a positive R² of 0.1698 for the unseen subset, that estimate came from only 53 records distributed across chemically diverse categories and was not evidence of reliable extrapolation to arbitrary new materials.

The model pipeline could technically process a previously unseen categorical level because its one-hot encoder was configured to accept unknown categories. That software capability did not establish scientifically valid chemical extrapolation, however. The analysis instead supports interpolation and screening within the chemical and experimental domain represented in the training data. The authors noted that molecular descriptors, chemical fingerprints, learned material embeddings and additional biological, genomic or pharmacokinetic information might help future systems extend beyond familiar core materials. Independent external datasets and prospective experiments would also be required to determine whether model-generated priorities correspond to better delivery outcomes in new laboratory or clinical settings.

Overall, the study presents explainable machine learning as a useful but bounded research tool for cancer nanomedicine. Its models can summarize patterns linking nanoparticle characteristics, biological context and administration conditions to tumor delivery measurements, and SHAP makes those learned relationships easier to inspect. At the same time, the conservative results show that algorithmic complexity cannot overcome heterogeneous evidence or the difficulty of transferring predictions between studies. Prediction for unseen nanoparticle identities was substantially harder than row-wise validation implied, performance depended on which formulations entered each partition, and cross-study transport was limited. The framework may therefore help prioritize candidate formulations and generate hypotheses for experimental work, but it should complement rather than replace laboratory validation, and its feature rankings should not be treated as universal design rules or causal mechanisms.

Subject of Research: Explainable machine-learning prediction of nanoparticle tumor delivery from the updated Nano-Tumor Database

Article Title: Predicting nanoparticle tumor delivery efficiency using explainable machine learning for cancer nanomedicine design

Article References: Karim, A., Sutikno, Syarif, A. M., & Jeong, I. C. (2026). Predicting nanoparticle tumor delivery efficiency using explainable machine learning for cancer nanomedicine design. Frontiers in Bioengineering and Biotechnology, 14, Article 1919132. https://doi.org/10.3389/fbioe.2026.1919132

Image Credits: AI Generated

DOI: 10.3389/fbioe.2026.1919132

Keywords: nanoparticle delivery, cancer nanomedicine, machine learning, explainable AI, SHAP, tumor biodistribution, model validation, nanoparticle materials, Predicting, nanoparticle, tumor, delivery