(W1230-11-75) The Application of Machine Learning in Predicting Bioink Printability

Purpose: 3D bioprinting, a paradigm shift in biofabrication, has been extensively applied in a variety of research areas, including drug discovery and delivery, disease modeling, tissue engineering, and regenerative medicine [1–4]. As an essential component of bioprinting, bioink provides a supportive environment for cell proliferation and differentiation, which is crucial for creating 3D structures with biological functionality. To formulate ideal bioinks, an important feature that should be accommodated is printability, the ability to generate predesigned 3D constructs with good shape fidelity and integrity [5,6]. Printability is governed by multiple factors, including bioink properties and printing parameters such as printing speed, printing pressure, temperature, nozzle characteristics, etc. [6,7]. Conventional optimization of bioink printability still strongly relies on the laborious trial-and-error approach, which is not efficient in finely tuning a myriad of process parameters. Moreover, many bioprinting studies reported only the optimized parameter values instead of fully utilizing undesirable experimental results to enhance the work efficiency for future researchers. Thus, there is an unmet need to build up a platform for bioink optimization, which combines the information from both desirable and undesirable results, so that researchers can use this platform to design and predict the properties of new bioinks developed under adjusted experimental conditions. Recently, machine learning (ML)/artificial intelligence (AI) is attracting tremendous research interest, which has emerged as a powerful tool to efficiently optimize process parameters and predict formulation properties [8]. The aim of this study is to employ ML modeling to develop a platform to evaluate bioink printability stepwise qualitatively and quantitatively. This platform is also expected to predict the printability of newly developed bioinks and to efficiently optimize multiple process parameters to achieve printed structures with good shape fidelity.

Methods: Sodium alginate (SA) and methylcellulose (MC) were employed to prepare hydrogels at different concentrations. In addition to bioink concentrations, other investigated process parameters included needle type (conical or cylindrical), needle size (22 or 25G), temperature (room temperature (RT) or 37 ), printing speed (1 or 2 mm/s), and printing pressure (10-200 kPa). The evaluation of printability is a two-step process. During the initial screening, filament formation and layer stacking test were performed to qualitatively evaluate printability. Different process parameter combinations that resulted in printable and unprintable bioinks were labeled as 1 and 0, respectively. Next, eight classification ML models including decision tree (DT), random forest (RF), light gradient boosting machine (LightGBM), extreme gradient boost (XGBoost), K-nearest neighbor (KNN), support vector machine (SVM), artificial neural network (ANN), and deep neural network (DNN) were constructed to predict the printability categories of different parameter combinations. The parameter combinations that let to printable bioinks were further employed to bioprint 3D constructs. Filament spreading ratios (FSRs) of printed structures were measured to quantitatively assess bioink printability. Moreover, eight regression ML models including DT, RF, LightGBM, XGBoost, KNN, ANN, DNN, and multiple linear regression (MLR), were established to predict the FSR values of initially screened parameter combinations. Subsequently, the most predictive classification model and regression model were employed to rank the importance of process parameters.

Results: The filament formation and layer stacking test indicated that increasing the concentration of SA in hybrid SA/MC bioinks improved the printability when the other experimental conditions were kept constant. Both tests were performed using 22G conical needle under RT. Based on the results, the 1% SA/3% MC bioink was liquid like and showed poor printability (Fig. 1Ai-Bi). The 1% SA/5% MC bioink showed certain degree of gelation and inadequate printability (Fig. 1Aii-Bii). On the other hand, 1% SA/9% MC bioink passed both tests and showed excellent printability (Fig. 1Aiii-Biii). Moreover, as is shown in Fig. 1C, both higher temperature and greater pressure led to increased filament spreading. Furthermore, higher printing speed could cause filament break at lower printing pressures but reduced the FSR at higher printing pressures. Additionally, the (receiver operating characteristic) ROC curves of different ML models and the confusion matrix of most predictive model, XGBoost, were shown in Fig. 2A and Fig. 2B, respectively. The feature importance ranked by XGBoost indicated that MC concentration was the most significant factor affecting the qualitatively evaluated printability Fig. 2C. Higher MC concentration improved bioink printability. For quantitatively evaluated printability, FSR, the data distribution and scatter plots of predicted values versus experimental values were shown in Fig 3A and Fig 3B, respectively. The feature importance ranked by LightGBM, the most predictive regression ML model, was shown in Fig 3C.

Conclusion: In this work, we successfully employed ML modeling to develop a platform to comprehensively evaluate bioink printability in a two-step process. This platform showed excellent predictive performance, which shows great promise in predicting the printability of newly developed bioinks and fast track optimizing multiple process parameters to achieve printed structures with good shape fidelity.

References: 1. Satpathy, A.; Datta, P.; Wu, Y.; Ayan, B.; Bayram, E.; Ozbolat, I.T. Developments with 3D bioprinting for novel drug discovery. Expert Opin. Drug Discov. 2018, 13, 1115–1129.
2. Chircov, C.; Grumezescu, A.M. Three-dimensional bioprinting in drug delivery. Mater. Biomed. Eng. 2019, 19–40.
3. Wang, Z.; Kapadia, W.; Li, C.; Lin, F.; Pereira, R.F.; Granja, P.L.; Sarmento, B.; Cui, W. Tissue-specific engineering: 3D bioprinting in regenerative medicine. J. Control. Release 2021, 329, 237–256.
4. Loukelis, K.; Helal, Z.A.; Mikos, A.G.; Chatzinikolaidou, M. Nanocomposite Bioprinting for Tissue Engineering Applications. Gels 2023, 9, 103.
5. Malekpour, A.; Chen, X. Printability and cell viability in extrusion-based bioprinting from experimental, computational, and machine learning views. J. Funct. Biomater. 2022, 13, 40.
6. Naghieh, S.; Chen, D. Printability--a key issue in extrusion-based bioprinting. J. Pharm. Anal. 2021.
7. Thayer, P.; Martinez, H.; Gatenholm, E. Manufacturing of Biomaterials via a 3D Printing Platform. 3D 4D Print. Biomed. Appl. Process Eng. Addit. Manuf. 2019, 81–111.
8. Wang, J.; Aghda, N.H.; Jiang, J.; Habib, A.M.; Ouyang, D.; Maniruzzaman, M. 3D bioprinted microparticles: Optimizing loading efficiency using advanced DoE technique and machine learning modeling. Int. J. Pharm. 2022, 628, 122302.

Acknowledgements: The authors gratefully acknowledge funding through the Food and Drug Administration, National Institutes of Health (1R01FD007456-01) to M. Maniruzzaman.

Fig 1. (A) Filament formation test. (i) Droplet formation. (ii) Gel accumulation around the tip. (iii) Filament formation. (B) Layer stacking test. (i) Lines of liquid-like bioink coalesce and show great spreading when printed onto the glass slide. (ii) Lines of gel-like unprintable bioink coalesce and show moderate spreading. (iii) Lines of gel-like printable bioink stack on top of one another. (C) Comparison of bioink printability under different temperatures (RT and 37℃), printing speed (1 mm/s and 2 mm/s), and printing pressure (40 kPa-70 kPa). The scale bar is 2mm.

Fig 2. (A) ROC curves of eight classification models. (B) Confusion matrix of XGBoost model. (C) XGBoost model interpretation by SHAP summary plot of process parameters.

Fig. 3 (A) Data distribution of FSR values. (B) Scatter plots of predicted values calculated by different regression ML models versus experimental values on the training and test subset. (C) LightGBM model interpretation by SHAP summary plot of process parameters.