Answers are grounded in this paper and cite their sources.
Introduction Literature-derived machine learning may guide preclinical cancer nanomedicine, but its value for predicting treatment efficacy depends on what it learns and where its predictions remain reliable. We examined quantitative predictive fidelity, the relative contributions of treatment context and nanoparticle design, and generalization across experimental domains. Methods We independently reanalysed the NanoAnalysis resource of 745 preclinical studies of inorganic nanoparticles in cancer. The released modelling matrix included 397 studies and 130 predictors. Repeated cross-validation assessed classification of >90% tumor-volume reduction relative to untreated controls and prediction of continuous reduction. Feature-family comparisons and grouped permutation analyses examined signal attribution. Last-author-grouped validation, within-corpus chronological tests and domain holdouts benchmarked against random holdouts matched for sample size and outcome-class counts assessed generalization. Results The full random forest achieved an out-of-fold area under the receiver operating characteristic curve (AUC) of 0.712 (95% bootstrap confidence interval 0.659–0.763), average precision of 0.688 and Brier score of 0.214. Continuous prediction explained little variation (R 2 = 0.118), with mean absolute error of 12.1 percentage points versus 12.9 for a training-fold median baseline. A 13-feature model comprising therapy modality and combination-therapy status achieved AUC 0.695, whereas nanoparticle-design-only features achieved 0.568. Removing treatment context reduced AUC to 0.607; removing nanoparticle-design features retained 0.686. Last-author grouping had little effect on discrimination, while the latest chronological test included only 30 studies and had wide uncertainty. Iron-oxide holdout AUC was 0.583 versus a matched-random median of 0.709, whereas gold showed little penalty. Discussion Moderate discrimination of high-response studies coexisted with limited prediction of response magnitude. Under the evaluated representation and models, treatment context carried most of the recoverable discriminatory signal, and transferability varied by material domain. For preclinical pharmacology, these findings support context-aware hypothesis generation and emphasize quantitative error and applicability-domain evaluation before using literature-derived predictions to select formulations.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: