Materials

Indomethacin (IND) was purchased from Acros Organics (Geel, Belgium). Transcutol® P (highly purified diethylene glycol monoethyl ether), Capryol® 90 (propylene glycol monocaprylate), Labrafac™ lipophile WL 1349 (triglycerides medium-chain), Labrafac™ PG (propylene glycol dicaprylate), Labrafil® M 2125 CS (linoleoyl macrogol-6 glycerides), Labrafil® M 1944 CS (oleoyl macrogol-6 glycerides), and Labrasol® ALF (caprylocaproyl macrogol-8 glycerides) were used as liquid lipids (LL). Compritol® 888 ATO (glyceryl behenate), and Precirol® ATO 5 (glyceryl distearate) were selected as solid lipids (SL). Both LL and SL were kindly provided by Gattefossé (Saint-Priest, France). Epikuron® 145 V (lecithin) (deoiled phosphatidyl choline-enriched lecithin) was a kind gift from Cargill (Wayzata, MN, USA). Polysorbate 80 (Tween® 80), dialysis membrane (Spectrum. Labs Spectra/Por, MWCO 3.5 kDa), octadecylamine (stearylamine) (SA) and the D- (+)-Mannose were acquired from Sigma Aldrich (St Louis, MO, USA). Sodium cyanoborohydride 95% and Anthranilic acid 98% were acquired from Acros Organics (Geel, Belgium). Acetate buffer at pH 4 was prepared using acetic acid 0.2 M from Sigma Aldrich (St Louis, MO, USA) and sodium acetate 0.2 M from Scharlab (Barcelona, Spain). Ultrapure water (Milli-Q® plus, Millipore Ibérica, Madrid, Spain) was used throughout all the experiments and the remaining solvents and reagents were analytical or HPLC grade.

Selection of liquid and solid lipids

To select the most appropriate liquid lipid (LL) for the indomethacin encapsulation (IND), a solubility study was conducted following a procedure described previously by Gaspar et al. with slight modifications. To this end, 200 mg of IND were added to 1 mL of each LL and stirred for 48 h at 300 rpm. Then samples were centrifuged at 12,000 rpm and 20 °C for 30 min to remove drug crystals [10]. Subsequently, samples were suitably diluted in acetonitrile, and the IND concentration was quantified spectrophotometrically (Agilent Technologies, Spain) at 322 nm. Solubility studies were carried out in triplicate.

The most suitable solid lipid (SL) was selected by evaluating the solubility of IND in Compritol® 888 ATO (melting range between 65 °C and 77 °C) and Precirol® ATO 5 (melting range between 50 °C and 60 °C). To accomplish this, aliquots of each SL, weighing 200 mg were heated in a water bath at 80 °C until fully melted. Gradually increased amounts of IND were added to each lipid until a precipitate of non-solubilized drug became visible. The solubility of IND in the SL is expressed as the maximum amount that can be solubilized without the presence of precipitate [10].

Finally, the miscibility of the selected LL and both SL was examined blending them in various proportions (75:25, 50:50, 25:75). The mixtures were then heated in a water bath at 80 °C for 5 min. The presence of phase separation upon observation indicated lack of miscibility.

NLC formulation and design

NLC were prepared using hot shear homogenization similar to previously reported [25]. In brief, a lipid phase (300 mg) comprising the selected lipids with or without indomethacin (Transcutol® P and Compritol® 888 ATO) was prepared. Simultaneously, a separate aqueous phase (10 mL) containing lecithin and Tween® 80 was also prepared. Both phases were obtained as specified in Table 1 and then heated in a water bath at 80 °C for 5 min. Subsequently, the aqueous phase was added to the lipid phase, and the mixture was homogenized for 10 min at 14,800 rpm using an Ultra-Turrax T25 (IKA Labortechnik, Staufen, Germany). The obtained dispersion was rapidly cooled in an ice bath with gentle agitation for 2 min. Each formulation was characterized before and after dialysis (MWCO 3.5 kDa) for the time specified in Table 1 to eliminate the non-incorporated components [24, 25]. A reduced experimental design for four variables (LL/SL ratio, Tween® 80 concentration, amount of lecithin, and dialysis time) with a minimum pattern of 3 was established using DataForm® v3.1 software (Intelligensys Ltd, UK). Both blank (unloaded) and IND loaded NLC formulations were prepared as indicated in Table 1 (Formulations N01 to N30). The amount of drug used for each loaded formulation was the maximum not leading to a drug precipitate. The obtained database (highlighted in grey) was modelled (Model A) and the obtained model was then validated (Formulation N31). In a second step, as described in Fig. 1, a new variable, stearylamine (SA), required for carbohydrate functionalization to endow the NPs with moieties able to covalently link these components, was incorporated [17, 30]. These new formulations were prepared adding SA to the lipid blend following the conditions shown in Table 1 (Formulations N32 to N44). The increased database was modelled again (Model B) (Formulations N01 to N44), the obtained model was re-validated (IND-NLC-SA), and the stability of the formulation was assessed after three months of storage. Finally, the stable formulation containing the functionalization linker was covalently functionalized with mannose though the Maillard reaction and characterized.

Fig. 1

Stepwise optimization procedure to obtain optimized mannose-functionalized nanostructured lipid carriers (NLC) using hybrid artificial intelligence (AI) tools

Table 1 Formulations prepared for Model A (grey) and Model B (white) along with the distinctive composition characteristics and purification conditionsMannose labelling and NLC functionalization

Mannose was labelled with anthranilic acid (2-aminobenzoic acid) (2AA) through a Maillard reaction similarly to already reported [11]. Briefly, 6 mg of anthranilic acid (2AA) was dissolved in 100 µL of dimethyl sulfoxide: acetic acid (7:3, v: v) containing 1 M of sodium cyanoborohydride (NaBH3CN). Then, mannose (25.2 mg) was dissolved in the previously prepared solution. The resulting mixture was vigorously vortexed at maximum speed (Fisherbrand, USA), protected from light, and incubated for 12 h at 37 °C. After this time, the labelled mannose was diluted in acetate buffer, purified employing a cyano-modified silica gel-SPE column (Chromabond®), and lyophilized.

Proton nuclear magnetic resonance spectroscopy analysis (1H-NMR) was conducted to confirm the labelling. For this purpose, fluorescently labelled mannose and unlabelled mannose were dissolved in deuterated water (D2O) obtained from Sigma Aldrich (St Louis, MO, USA) at 200 mM, while anthranilic acid was dissolved in deuterated dimethyl sulfoxide (DMSO-D6; Cambridge Isotope Laboratories, Inc. (Andover, MA, USA)) at the same concentration. The solutions were then placed in NMR tubes (600 µl) to be characterized using a Bruker NEO-750 equipment (Karlsruhe, Germany).

NLC functionalization

Drug loaded NLC suspensions containing SA (IND-NLC-SA) were incubated with 50 mM fluorescently labelled D-(+)-mannose solution (MAN) in acetate buffer at pH 4. The mixtures were kept under magnetic stirring at 300 rpm (Fisherbrand, USA) protected from light at different timepoints. Subsequently, to eliminate unreacted components, formulations underwent a 12-hour dialysis using a cellulose membrane (MWCO: 3.5 kDa) under stirring in Milli-Q® water.

Various molar ratios of SA: MAN (1:2, 1:3, and 1:5) and incubation times (24 h, 48 h, and 72 h) were tested to select the best functionalization strategy. To assess the effectiveness of the functionalization, the non-linked MAN was quantified. For this purpose, formulations underwent ultrafiltration using Amicon® 100 kDa ultrafilters (Sigma Aldrich) at 10,000 g and 4 °C for 15 min. Subnatants were collected, and fluorescence was measured at 360 nm/425 nm using a plate reader (FLUOstar Omega, BMG Labtech, Germany). A calibration curve with MAN was also included during the fluorescence determination.

NLC characterization

NLC were characterized in terms of particle size (size), polydispersity index (PDI), and surface charge (ZP) using a Zetasizer Pro (Malvern Instruments, Malvern, UK). Samples were diluted with Milli-Q® water (1:10), placed in a specific cuvette (DTS 1070), and measured in triplicate using the parameters automatically selected by the software. All the measurements were performed at 25 °C ± 1 °C.

Additionally, the morphology of IND-NLC-SA nanoparticles functionalized with D-(+)-mannose at 1:2 SA: MAN ratio for 72 h (IND-NLC-SA-MAN) was evaluated through Transmission Electron Microscopy (TEM). Colloidal dispersions were placed onto cooper grids coated with a carbon membrane and stained with 2% (w/v) phosphotungstic acid solution for 2 min, following a previously described procedure [26]. The observation was conducted using a JEOL microscope (JEM 2010, Tokyo, Japan) equipped with a Gatan OriusTM camera (Gatan, Inc., Pleasanton, CA, USA).

Drug loading (DL) within the lipid network after purification (dialysis) of each formulation was calculated using Eq. 1, before and after storage. To determine actual drug content, samples were dissolved in acetonitrile and centrifuged at 12,000 rpm and 4 °C for 30 min. The supernatants were filtered through 0.45 μm and properly diluted in acetonitrile. IND quantification was performed spectrophotometrically at 322 nm.

$$DL \left(\%\right)=\left[\frac\right] x 100$$

(1)

NLC modeling and optimization

The NLC characterization data of the first-step formulations (Table S1; N01-N30) was modelled using FormRules®v4.03 (Intelligensys Ltd., UK) (Model A). FormRules® is a commercial software that combines two AI techniques, artificial neural networks (ANN) and fuzzy logic. This software uses the Adaptive Spline Modelling of Data (ASMOD) algorithm to generate models that relate inputs (ingredients and operation conditions) and outputs (formulation properties). Its ability to generate IF-THEN linguistic rules makes easier to understand the effects of the different ingredients and operating conditions on the process or products produced. This allows generating knowledge about the key variables to obtain IND-loaded NLC. FormRules® enables the use of different fitting statistical criteria. Among them, Structural Risk Minimization (SRM) was selected as it provided the highest predictability along with the simplest and most intelligible rules. Five variables were introduced as inputs: LL/SL ratio, Tween® 80 concentration, amount of lecithin, dialysis time, and the amount of IND. On the other hand, four parameters were included as outputs: Size, PDI, ZP, and DL.

The training parameters used by FormRules® software were the following: ridge regression factor of 1.0 e− 6, number of set densities: 2, set densities: 2 and 3, maximum inputs per submodel: 2, maximum nodes per input: 15, adapt nodes: true, C1 values over 0.75.

The determination coefficient of the Training set (R2) (Eq. 2) was used to establish the predictability of the models, where yi is the actual value in the data set, yi´ is the value calculated by the model, and yi” is the mean of the dependent variable. The greater the value of the training set R2, the higher the predictability of the model.

\( ^= \left[1-\frac^ (_-_)}^}^(_-_)}^}\right]x 100\%\) (2)

The comparison between predicted and experimental values has been conducted using an ANOVA. The absence of differences between both sets indicates the accuracy of the model.

On the other hand, INForm®v5.01 software (Intelligensys Ltd, United Kingdom), integrates two AI technologies, ANN and Genetic Algorithms (GA), and was specifically designed for optimizing pharmaceutical formulations. INForm® was used to obtain an optimized formulation using the initial dataset (Formulation N31). The training parameters used for the models were: 1 hidden layer, 2 nodes, transfer type (Asymmetric Sigmoid), output transfer type (Asymmetric Sigmoid), back propagation type (RPROP), target interactions (2000), target mean squared error (0.001), and random seed (10,000) and the optimization parameters were: number of populations = 1, number of interactions = 100, population size = 100, percentage of replacement = 50, mutation standard deviation = 0.1, and random seed = 1.

A second model (Model B) was developed using the data from formulations with and without the functionalization linker (Table S1; N01-N44). FormRules® was employed to model the database under the same conditions as previously described, adding the amount of SA (mg) as an additional input.

Based on these models, a protocol (composition and appropriate purification conditions) for suitable and stable NLC including the functionalization linker was established. Specifically, the protocol aimed at obtaining NLC with small particle size (< 120 nm), maximum surface charge, and the highest drug loading (> 2%) not showing differences in their physicochemical properties after three months of storage. For this purpose, INForm® was again employed. The training parameters used for the models were: 1 hidden layer, 3 nodes, transfer type (Asymmetric Sigmoid), output transfer type (linear), Back propagation type (RPROP), target interactions (1000), target mean squared error (0.0001), and random seed (10,000).

Statistical analysis

The obtained data are expressed as mean ± SD and analyzed using GraphPad Prism 8 software. The groups were compared by performing one-way analysis of variance (ANOVA) followed by post hoc Tukey’s Multiple Comparison Test. The confident interval was 95%.