(160bc) Machine Assisted Experimentation for Extrusion-Based Bioprinting

Motivation: Extrusion-based bioprinting (EBB) involves the deposition of biomaterials through a container using pressure exerted pneumatically or mechanically. When live cells are embedded within the biomaterials used, a combination of material parameters and printer settings impact its printability and cell viability when extruded, including nozzle outlet diameter, material concentration, and operating temperature. Thus far, the optimization of EBB parameters has been conducted through systematic wet-lab experimentation. This process can be laborious, and the information gained can be hard to translate towards different biomaterials and printers. We hypothesize that broadly applicable EBB design rules can be identified by applying machine learning (ML) to data available in the literature.

Approach: In this project, the objective is to create a cell viability and printability prediction model by training machine learning algorithms on a dataset of material concentration, printing settings, cell viability, and printability results accrued from 47 EBB manuscripts over the past 13 years. Two modeling approaches were used: regression and classification. Regression models compared include random forest regression, linear regression, support vector regression, and decision tree regression. Classification models compared include random forest classification, logistic regression, support vector classification, and decision tree classification. Coefficients of determination and mean squared errors were compared amongst regression models across different training data sizes to determine model performance. Prediction accuracy across different training data sizes was used to elicit better performing models than others for classification cases. Feature importance and collinearity analysis were conducted to understand what equipment, material, and cellular parameters impacted model behavior the most. A set of parameter combinations based on equipment and material availability in lab were then provided for chosen models to predict cell viability and printability outcomes. Decision trees from random forest and decision tree models were also generated to compare and supplement predicted values of parameter combinations used for wet-lab validation of models.

Results/Conclusions: Random forest regression models for cell viability and filament diameter predictions both elicited higher coefficients of determination while minimizing average mean squared error. Random forest classification models elicited higher prediction accuracy than other models tested. Feature importance testing based on decision trees generated from random forest and decision tree models indicated relatively major effects from specific material concentration, printing temperature, nozzle diameter, and printing pressure for cell viability predictions. Meanwhile, the nozzle diameter was by the far the most important feature affecting filament diameter model prediction. The relative importance of mentioned parameters correlates with effects on cell viability and filament diameter seen from previous studies in the EBB field.

Framing the prediction models as classification models demonstrated that relatively higher prediction accuracy on test set data (average prediction accuracy range of 77 to 75% for random forest and decision tree classification models respectively) can be achieved for cell viability outcomes. Relatively higher coefficients of determination scores (0.86) for random forest regression modeling of filament diameters indicate a relatively linear model based on information provided in published papers. Ongoing work involves validating prediction results from testable parameter combinations through wet-lab experiments.