Artificial intelligence-guided design of lipid nanoparticles for pulmonary gene therapy

Abstract

Ionizable lipids are a key component of lipid nanoparticles, the leading nonviral messenger RNA delivery technology. Here, to advance the identification of ionizable lipids beyond current methods, which rely on experimental screening and/or rational design, we introduce lipid optimization using neural networks, a deep-learning strategy for ionizable lipid design. We created a dataset of >9,000 lipid nanoparticle activity measurements and used it to train a directed message-passing neural network for prediction of nucleic acid delivery with diverse lipid structures. Lipid optimization using neural networks predicted RNA delivery in vitro and in vivo and extrapolated to structures divergent from the training set. We evaluated 1.6 million lipids in silico and identified two structures, FO-32 and FO-35, with local mRNA delivery to the mouse muscle and nasal mucosa. FO-32 matched the state of the art for nebulized mRNA delivery to the mouse lung, and both FO-32 and FO-35 efficiently delivered mRNA to ferret lungs. Overall, this work shows the utility of deep learning for improving nanoparticle delivery.

Figure 1. Workflow for LiON, deep learning-based LNP design.

A) Collect available LNP data, typically collected from a screen generated via combinatorial synthesis. B) Train predictive MPNN-based DL model on training data. Inputs are the chemical structure of the ionizable lipid along with metadata (formulation details, cargo, target of LNP). C) Rank lipids from novel library using trained model. D) Screen top candidates in vivo. E) Follow up and structurally optimize individual promising LNPs. F, G) Performance of model as measured by correlation between predicted and experimental delivery results on held-out test set using F) a random training test split and G) an amine-based training-test split. Only datasets with n>10 in test set are shown. p values for significance of correlation used beta distribution for null hypothesis of no correlation and were two-sided.

Figure 2.. LiON for optimizing IV liver delivery of branched ester-based library.

A) Michael addition chemistry for generation of branched-tail ionizable lipids. B-F) FFL mRNA delivery testing to the liver, 1ug dose delivered IV, imaging after 6 h. B) Screen of top DL candidates with EPA (T08) and DHA (T16) tails. n=3 for RM-2–133-3 control, n=1–2 for rest of screen. C) RJ-A03-T16 has improved liver delivery compared to RM-133–3 (which according to this work’s nomenclature would be designated RJ-A03-T08). n=4. Unpaired t test, t=3.747, d.f.=6. D) Screen of top DL candidates with ALA (T30) and GLA (T34) tails (n=2 for RM-133–3, n=1 otherwise). E) Evaluation of top candidates RJ-A14-T30 and RJ-A14-T34 relative to controls (n=3–4 for controls, n=7 for RJ-A14-T30, n=4 for RJ-A14-T34). F) Screening of candidates with tail T01, with RJ-A30-T01 as top candidate (n=3 for RJ-A14-T34, n=1 otherwise). G) Testing of top candidates relative to controls shows RJ-A30-T01 appears to be the top performer (n=4).. H) Serum Epo 6 h after Epo mRNA delivery (n=4). All doses for this figure are 1ug. Error bars, mean +/− SEM. E, G, H) One-way ANOVAs with Šídák’s multiple comparisons test between controls (black) and novel lipids (blue).

Figure 3.. Combining LiON with 4-component reaction chemistry.

A) 4-component reaction scheme to generate novel ionizable lipids. B-K) IM FFL mRNA delivery testing, imaged 6 h after injection in BALB/C mice unless otherwise indicated. B) Initial IM delivery testing with ML3 and ML5 (1ug dose, BL/6 mice) compared to cKK-E12 (n=1 mouse/LNP injected on each flank). C) Structures of ML3 and ML3Me. D) delivery of ML3Me is better than delivery of ML3 (1ug dose, n=1 mouse/LNP injected on each flank). E) Testing of top DL candidates against SM-102 and cKK-E12 controls (5ug dose, n=3) shows promise for some candidates. F) IM delivery of FO-32 compared to cKK-E12 and SM-102 (1ug dose, n=4). G) Structures of top candidates FO-32, FO-35, and EB-66. H-K) Comparison of delivery of formulation-optimized DL-designed lipids compared to cKK-E12 (formulation-matched and with original KK formulation) and SM-102 (clinical vaccine formulation), at H) 0.01 (n=4), I) 0.1 (n=3), J) 1 (n=5–6), and K) 10ug (n=3, n=2 for SM-102 due to missed IP injection) doses. Error bars, mean +/− SEM. H-K: One-way ANOVAs with Šídák’s multiple comparisons test between controls (cKK-E12 KK and F3, SM-102) and novel lipids.

Figure 4.. Respiratory tract delivery.

A) IN FFL delivery (1.5ug dose, n=4) comparison between lead DL candidates (FO-32, FO-35, EB-66) and controls (cKK-E12, SM-102, Lipid 331). B) OPA FFL delivery comparison (1ug dose, n=8 for novel LNPs, n=4 for controls) between lead LiON candidates and controls. A,B) One-way ANOVAs with Šídák’s multiple comparisons test between controls (cKK-E12, SM-102, Lipid 331, RCB-4–8) and novel lipids. C,D,E) OPA FFL delivery (1ug dose, n=10, two outliers removed for cKK-E12 due to missed IP injections, two-sided Welch’s t test) for FO-32 compared to top control cKK-E12. C) Lung luminescence, D) Tracheal luminescence quantified following background subtraction due to low signal, E) Representative lung image. F) Nebulized FFL delivery (500ug dose, n=5, 2-way ANOVA with Šídák’s comparison test for IR-117–17 vs FO-32 in each tissue) for FO-32 compared to top control IR-117–17 for nose, lung, and trachea. IN delivery is in BALB/C, all OPA and nebulized testing in BL/6 mice. All error bars, mean +/− SEM. Outlier removal based on ROUT method (Q=0.05) and does not change any statistical significance. Data with outliers can be found in Source Data.

Figure 5.. LNP delivery testing in ferret model.

A) Ferret Cre reporter model; Cre activity converts cells from mtdTomato to mEGFP expression. B) Timeline of LNP-Cre delivery experiments. C-E) Cre mRNA delivery to adult (>5 months old) ferret lungs (1 ferret per LNP, 0.6mg/kg total mRNA delivered in 2mL/kg volume). Detailed manually selected images of highly transfected airways showing transfection of (C) IR-117–17, (D) FO-32, and (E) FO-35. For FO-35-treated ferret, left upper lobe zoomed-in image shows transfection at sub-mucosal glands. Blue arrows indicate apical side of trachea; zoomed-out trachea for 117–17 shows undetectable transfection in entire section.