Storage stability of lipid nanoparticles: a curated benchmark dataset, systematic threshold analysis, and per-study...

Lipid nanoparticles have emerged as the preferred delivery vehicle for nonviral nucleic acid therapies. Nonetheless, the key factors that govern their stability during storage remain poorly understood. Design-of-experiment studies have identified storage duration, temperature, and cryoprotectant type as critical variables; however, the applicability of these findings across formulation procedures, lipid compositions, and biological delivery systems has not been thoroughly tested. We address this gap by conducting a systematic analysis of 613 LNP formulations from 18 independent published studies. A systematic, pooled threshold analysis of storage duration and temperature effects, as well as a rigorous test to determine whether per-study machine learning models can predict storage-induced instability based on formulation and storage parameters. After using nested, formulation-grouped cross-validation to eliminate repeated-measures leakage in naive train/test splits, we found that particle-size-change models (Model 1) showed no reliable predictive validity in any of the six adequately powered studies (median R² = -1.53 across 12 eligible studies; 0/12 studies reached R² ≥ 0.70). Biological-activity-change models (Model 2) succeeded in one of two adequately powered studies (P10, R2 = 0.854) but failed in the other (P3, R² = -0.556), indicating that predictive success is study-specific rather than a generalizable property of the modeling approach. Cross-study prediction training and testing on held-out studies failed uniformly across four algorithms tested (grouped cross-validation mean R² between -1.73 and -0.27). The largest study dominates the pooled cross-study R² values, highlighting the need for per-study reporting. A pooled threshold analysis of all 613 formulations, which was not influenced by the modeling limitations specific to each study, revealed that lyophilization is protective only after 90 days, which is the critical storage duration beyond which median biological activity retention drops below 80%. These findings show that per-study machine learning models are not yet a reliable tool for predicting LNP storage instability using currently reported data and standard molecular and formulation descriptors, and that the main obstacle to advancement in this field is standardized, higher-resolution data reporting rather than additional modeling.