Fig. 3: LNPDB facilitates an improved deep learning model for predicting LNP delivery performance.

a Our deep learning model for predicting LNP delivery performance—lipid optimization using neural networks (LiON)—is improved for 5 of the 7 studies evaluated when trained on the LNPDB compared to the original smaller dataset used in our prior study¹⁷. The plot illustrates the performance of LiON as measured by Spearman correlation between predicted and experimental delivery results on test datasets using an amine-based 70%-15%-15% train-validation-test split. Datasets shared between LNPDB and the original dataset are compared. b LiON trained on LNPDB significantly outperforms AGILE¹⁶, an alternative deep learning model for predicting LNP delivery performance, for the 4 fully held-out datasets evaluated. Bars denote mean Spearman correlation coefficient values, and error bars denote ±standard deviation across five train/validation splits with a fixed held-out test set. p values resulting from a two-tailed Student’s t-test are shown. Datasets evaluated are KW_2014³⁰; KZ_2016³²; LM_2019⁷; SL_2020³³; SL_2021³⁴; BL_2023³⁶; BL_2024¹⁵, ZC_2023⁴². Comparable results measured with Pearson correlation are shown in Supplementary Fig. 3. Source data are provided as a Source data file.