A hydrophilic 3D flow focusing microfluidic glass chip with 100 μm channels (3,200,433) was purchased from Dolomite Microfluidics (United Kingdom) and used to create blank or IND-loaded PLGA droplets. Indomethacin, PLGA (lactide:glycolide 50:50), polyvinyl acetate (PVA) (MW 9,000–10,000, 80% hydrolyzed), and Dichloromethane (DCM) were supplied by Sigma Aldrich (United Kingdom). To monitor droplets generation, a digital microscope (Dolomite, United Kingdom) was used. Fluids were injected and controlled into the microfluidic device by a flow control system (Fluigent, France).

Blank and IND-loaded PLGA MPs were manufactured with a 3D flow focusing microfluidic device. Initially, to generate blank PLGA droplets, different PLGA solutions were prepared at different concentrations (1, 2, 5% w/w). For each solution, an appropriate amount of PLGA was dissolved in DCM with continuous mechanical stirring to completely dissolve the polymer. To prepare droplets, 1% w/v PVA was used as an aqueous continuous phase and injected separately into two microfluidic inlets, whereas PLGA in DCM was the dispersed phase and injected into the central inlet. To evaluate the impact of different microfluidic production parameters, continuous and dispersed phases were injected at different flow rates. After optimizing the droplet generation conditions, IND-loaded PLGA droplets were prepared similarly to the blank ones, except the dispersed phase was DCM containing 2% PLGA and 0.5% IND. The generated droplets were collected in PVA aqueous solution to prevent droplet coalescence. Evaporation of DCM and generation of MPs were done using a rotary evaporator under reduced pressure at RT. Solidified blank or IND-loaded PLGA MPs were collected by centrifugation at 1,500 rpm for 5 min, and rinsed with deionized water to remove excess PVA.

The multilayer perceptron (MLP) ANN was employed using Statistica, version 13.3 software (TIBCO Sofware Inc, 2017). The ANN model was trained using the experimental data for the generation of blank PLGA droplets. A neural network consisting of an input, hidden, and output layer, with 3-6-1 neurons, respectively, was found to be adequate for predicting PLGA droplet size. The parameters chosen as inputs for the ANN were simple, including PLGA concentration and the flow rates of both PLGA and aqueous phases. The output layer consisted of the measured PLGA droplet sizes. All input features are continuous and were normalized to the range 0–1. Learning parameters were varied to optimize training performance and prediction accuracy. Tanh and Identity were used as the hidden and the output activation functions, respectively. The total dataset consisted of 23 cases, which has been further divided into ∼60, 20, and 20% for the training, test, and validation sets (Table 1).

Droplet/particle size distribution and morphology was examined using microscopic image analysis. The average particle size and size distribution were analyzed using ImageJ Software Version1.52a (NIH, US). Distribution is expressed as the polydispersity index (PDI), and was calculated with the following formula: P D I   % = s t a n d a r d   d e v i a t i o n a v e r a g e   d r o p l e t   d i a m e t e r   × 100

A low PDI value indicates a narrow size distribution (monodisperse particles), while a high PDI value indicates a wide size distribution (polydisperse particles). The size distribution is reflected narrow for a PDI value < 5%.

The IND loading was determined by dissolving 10 mg IND-loaded PLGA MPs in DMSO, and then diluting with PBS (pH 7.4). The drug content of the MPs was quantified by measuring the UV absorbance at 320 nm, and then the results were compared to a standard curve of known concentrations of IND (Supplementary Figure S1). Drug loading (DL%) and encapsulation efficiency (EE%) were calculated using the following formulas: D L   % =   mass of drug in MPs t o t a l   m a s s   o f   M P s   × 100 E E   % =   mass of drug in MPs i n i t i a l   m a s s   o f   d r u g   × 100

UV spectroscopy was used to monitor the stability of the produced MPs, in terms of IND release. The release behavior of 10 mg IND-loaded PLGA MPs was determined by dialysis against 10 ml PBS (pH 7.4) at 37°C, with continuous agitation. At predetermined time intervals, 100 μL of PBS was withdrawn and replaced with the same amount of fresh PBS to maintain the dissolution medium at a constant volume. The amount of released IND was determined by UV spectrophotometry.

Generation of monodisperse particles is a challenging task that arises the impact of particle sizes on the drug release kinetics. Thus, microfluidic devices are attracting more attention due to their ability to control the physical properties of generated particles, including size, dispersity, and shape. In the current study, highly uniform PLGA droplets, either blank or loaded with IND, were synthesized using a 3D flowfocusing microfluidic chip. Figure 1 shows the process of generation and solidification of PLGA droplets with or without IND, monitored by optical microscopy at different time intervals. Initially, prior to drug loading, engineering of blank PLGA droplets was investigated both experimentally and computationally to assess the effect of PLGA concentration and microfluidic flow rates of the continuous and disperse phases on generation of monodisperse droplets with a low PDI value.

Machine learning is a popular AI technique. ANNs, in particular, are commonly used machine learning technique used in pharmaceutical applications because of its powerful ability to model nonlinear relationships (Damiati, 2020). In our study, experimental data from the generation of blank PLGA droplet sizes were used as the target output to train an ANN, with input data furnished by the corresponding PLGA concentration, and PLGA and PVA flow rates (Figure 3A). The developed ANN offered highly accurate predictions of PLGA droplet sizes with residuals randomly scattered in the range ±5 µm (Figure 3B). The correlation of the observed and predicted droplet size data of the whole dataset was high (r ² = 0.990). The correlations of the predicted and observed droplet size of the training, test and validation datasets were 0.992, 0.997, and 0.990, respectively. The sensitivity analysis results of the trained ANN model provided key information on the relative importance of the input parameters in defining droplet sizes. By this means, it turned out that the order of importance of the input parameters was: PLGA flow rate > aqueous phase flow rate > PLGA concentration. Hence, the obtained data from the in silico and experimental protocol for PLGA MPs synthesis can be used further to generate drug-loaded particles at a desired size by easy, quick, and economical means.

Synthesis of IND-loaded PLGA droplets were fabricated utilizing the same 3D flow-focusing system described above for blank droplets. On the basis of the obtained results from prediction and experimental generation of the blank PLGA droplets, we chose a concentration of 2% (w/v) PLGA and flow rates of 3.7 and 9.2 µL/min for the disperse and continuous phases, respectively, as optimal conditions to generate size-tunable IND-loaded droplets. The chosen conditions were able to generate blank PLGA droplets at the size of 38.82 ± 0.5 µm, and IND-loaded PLGA droplets at size of 45.35 ± 0.4 µm (Figure 4). However, generation of IND-loaded PLGA droplets was not affected by adding IND to the disperse phase, but significantly increased droplet size by around 16.82%. Upon DMC evaporation, the droplets were solidified and shrunk by 30 and 45% for the initial blank and loaded PLGA diameters, respectively. Slow evaporation of the organic solvent from droplets resulted in complete annealing of the polymer which improves particle stability and leads to slower degradation rates (Allison 2008). Converting droplets to particles led to a significant increase in the PDI values, but still all formulations were monodisperse with low PDI values (<5%). A narrow size distribution confirms generation of a homogeneous droplet/particle size population.

DL and EE are two parameters considered crucial parameters to assess the properties of drug-loaded MPs and drug cumulative release profiles. Generated IND-loaded MPs with an average size of 45.35 µm showed an IND loading content of 7.79%, while the EE% was 62.35%. A possible explanation for these results is that IND-loaded PLGA droplets have a relatively small size, which means a high surface:volume ratio. Before solidification and formation of IND-loaded PLGA MPs, loos of the drug occurs at the droplet surface, and leaking to the external aqueous phase leads to a low DL%. Once the polymer droplets solidify, IND is entrapped in the PLGA matrix and this stops drug loss. Similar results were obtained by Chen et al. (2017) for Gefitinib-loaded PLGA microspheres with different size-fractions. In general, high EE reflects the affinity between IND and PLGA, and the hydrophobic nature of IND which increases its encapsulation into polymer particles (Tomoda et al., 2012). However, an improved DL% can be achieved by increasing the initial drug loading, but this may increase particle size, whereas increasing EE% can be achieved by lowering the polymer concentration within the formulation (Roces et al., 2020). Furthermore, when the solubility parameters of the drug and polymer are close, it increases compatibility between them and cause loading of more of the drug. Here, the solubility parameters of PLGA and IND are 28 and 24 MPa, respectively, which enhances drug loading into the PLGA MPs (Liu et al., 2005).

The release behavior of IND from PLGA MPs was evaluated in an experimental environment of PBS at pH 7.4 and at 37°C. The time dependence of the percentage of cumulative drug release from the IND-loaded PLGA MPs is shown in Figure 5. The drug release pattern followed biphasic drug release kinetics. The initial burst phase was up to 36% within 6 h, followed by accumulative release of >80% after 9 days. The delayed release may be attributed to slow diffusion of IND entrapped within the core of PLGA MPs into the dissolution medium.

The composition of a polymer governs the hydrophilicity and rate of degradation of a polymeric carrier system. Increasing the glycolic acid percentage in the oligomers accelerates the weight loss of a polymer. Compared to PLGA 75:25, PLGA 50:50 has a faster degradation that may be attributed to higher hydrophilicity and preferential degradation of the glycolic acid proportion. Moreover, PLGA 50:50 is more amorphous in nature (Makadia and Siegel, 2011; Tsai et al., 2013). All these factors may contribute to faster drug release from PLGA 50:50 particles. However, the loading amount of a drug is significantly influenced by the cumulative release profile. After a large initial burst, a high loading amount leads to a slow and pseudo-linear release. When more drug is distributed near the surface area of the MPs, a higher initial release and faster release rate are achieved. In contrast, a low drug-loading amount follows a gradual increase in release as a function of time (Liu et al., 2006). However, controlling the drug release rate from biodegradable polymers occurs via numerous mechanisms. In the case of the matrix structure, drug release depends on desorption of the surface-adsorbed drug, diffusion of the drug through a polymeric matrix, polymer matrix erosion, and a combination of erosion and diffusion processes (Soppimath et al., 2001).