From prediction to design rules: explainable machine learning analysis of cryoprotectant effects in GelMA-based cryobioprinting bioinks

Abstract

Abstract Machine learning has been used to predict cryoprotective-agent (CPA) effects in cryobioprinting, but prediction accuracy alone does not yield experimentally usable formulation rules. Here, we performed a secondary explainable analysis of Qiao et al ’s GelMA-HUVEC dataset using formulation-grouped validation, SHAP interpretation, stability metrics, and multi-objective ranking. A compact Random Forest model using permeable CPA identity, sugar CPA identity, permeable CPA concentration, sugar concentration, and storage day achieved MAE = 4.842%, RMSE = 6.106%, and R 2 = 0.649, with bootstrap 95% confidence intervals of 4.418–5.274% for MAE and 0.585–0.710 for R 2 . SHAP analysis identified permeable CPA concentration and storage day as the strongest viability drivers, with directional patterns favoring balanced 15% permeable CPA and 15% sugar CPA rather than simple concentration maximization. Multi-objective ranking identified GL_15_RA_15 as the strongest individual formulation, whereas EG + trehalose showed the highest average CPA-family performance. Cross-domain tests showed limited transfer between EG and GL formulations, indicating that the extracted rules should be treated as local, data-supported design hypotheses rather than universal cryoprotectant principles. Together, these results show that explainable secondary analysis can convert small cryobioprinting datasets into experimentally testable CPA hypotheses while explicitly defining their transferability limits.