The use of machine learning to predict the effects of cryoprotective agents on the GelMA-based bioinks used in extrusion cryobioprinting

• Published in: Bio-design and manufacturing 2023-07, Vol.6 (4), p.464-477
• Main Authors: Qiao, Qian, Zhang, Xiang, Yan, Zhenhao, Hou, Chuanyu, Zhang, Juanli, He, Yong, Zhao, Na, Yan, Shujie, Gong, Youping, Li, Qian
• Format: Article
• Language: English
• Subjects: 3-D printers; Artificial intelligence; Biomaterials; Biomedical Engineering and Bioengineering; Cells; Cryopreservation; Datasets; Dimethyl sulfoxide; Engineering; Ethylene glycol; Experiments; Gelatin; Gene amplification; Glycerol; Lactose; Learning algorithms; Machine learning; Mechanical Engineering; Mechanical properties; Neural networks; Polyethylene glycol; Research Article; Toxicity
• ISSN: 2096-5524; 2522-8552
• DOI: 10.1007/s42242-023-00244-4

Cryobioprinting has tremendous potential to solve problems to do with lack of shelf availability in traditional bioprinting by combining extrusion bioprinting and cryopreservation. In order to ensure the viability of cells in the frozen state and avoid the possible toxicity of dimethyl sulfoxide (DMSO), DMSO-free bioink design is critical for achieving successful cryobioprinting. A nontoxic gelatin methacryloyl-based bioink used in cryobioprinting is composed of cryoprotective agents (CPAs) and a buffer solution. The selection and ratio of CPAs in the bioink directly affect the survival of cells in the frozen state. However, the development of universal and efficient cryoprotective bioinks requires extensive experimentation. We first compared two commonly used CPA formulations via experiments in this study. Results show that the effect of using ethylene glycol as the permeable CPA was 6.07% better than that of glycerol. Two datasets were obtained and four machine-learning models were established to predict experimental outcomes. The predictive powers of multiple linear regression (MLR), decision tree (DT), random forest (RF), and artificial neural network (ANN) approaches were compared, suggesting an order of ANN>RF>DT>MLR. The final selected ANN model was then applied to another dataset. Results reveal that this machine-learning method can accurately predict the effects of cryoprotective bioinks composed of different CPAs. Outcomes also suggest that the formulations presented here have universality. Our findings are likely to greatly accelerate research and development on the use of bioinks for cryobioprinting. Graphic abstract