Evgeniya Borkovskaya1, Katherine Vilinski-Mazur2, Dmitry Kolomenskiy3 and Bogdan Kirillov4*
1Moscow Institute of Physics and Technology, Moscow, Russia
2Sechenov First Moscow State Medical University, Moscow, Russia
3Skolkovo Institute of Science and Technology, Moscow, Russia
4Sechenov First Moscow State Medical University
bogdan.kirillov [at] skoltech.ru
Abstract
3D bioprinters fabricate functional tissue constructs using cell-laden bioinks made out of composite hydrogels with tunable physical properties. The fidelity of resulting prints is governed by a multidimensional parameter space (e.g. concentration of each component, temperature, extrusion multiplier, speed and layer height). Grid search for optimal parameters is impractical and wasteful of both material and instrument time, however, most current approaches to address the parameter search task amount to empirical tuning based on the practitioner’s experience. A well-established field of active machine learning shows how to build algorithms to efficiently traverse the search space for new experimental parameters by controlling the model training with intelligent iterative selection and labeling of new data. We propose the application of Active Learning to accelerate the optimization of printing parameters for composite bioinks. A robust body of existing research in Active Learning shows reductions in dataset size requirements ranging from 50% to over 90% in comparable applications for computational metamaterial design, drug discovery and protein engineering. In our method, the Gaussian Processes surrogate model approximates the connection between printing parameters and geometric fidelity while a two-stage experimental design algorithm selects the next most informative experiment, maximizing data utility and minimizing trial costs. Ongoing empirical study performed using a custom low-cost extrusion bioprinter, diverse kappa-carrageenan-based compounds and deliberately suboptimal starting parameters demonstrates improvements in fidelity of the extruded lines during an iterative active learning loop. Our method is straightforward to adapt for new formulae of composite hydrogels and offers a practical, extensible workflow for tissue engineering applications.
Keywords: active machine learning, 3d bioprinting
Acknowledgement: The study was supported by grant No. 25-45-02119 from the Russian Science Foundation, https://rscf.ru/project/25-45-02119/.