Data-driven predictive design of engineered living hydrogels

TabPFN accurately predicts macroscopic material properties of living hydrogels from genetic and process parameters, with the strongest prediction for storage modulus G' at R2 = 85.1% on an independent validation set. Rational design of engineered living materials has been limited by the lack of quantitative relationships linking design parameters to material properties; this work establishes such a predictive mapping using a tabular foundation model under small-data conditions. Evaluated on an independent validation set, reporting R2 = 85.1% and a 48.0% RMSE reduction compared with linear regression.

Models were evaluated for predicting four properties: storage modulus G', fibre content, thickness, and permeability. Extends prediction from a single property to multiple macroscopic properties, covering E. coli-produced living hydrogels containing CsgA-based fibres fused to genetically encoded PEG-like biopolymers. All four properties were evaluated on the independent validation set, with G' showing the strongest prediction.

Property-guided design enabled identification of parameters for achieving living hydrogels with desired properties. Moves from prediction to inverse design, allowing designers to search genetic and process parameters for desired properties rather than relying solely on experimental screening. The text reports that property-guided design identified parameters for achieving desired properties, demonstrating the framework's capability.

The results apply to E. coli-produced living hydrogels containing CsgA-based fibres fused to genetically encoded PEG-like biopolymers, with predictions for four macroscopic properties: storage modulus G', fibre content, thickness, and permeability. The setting is research and engineering where a small amount of experimental data is used to build property predictions and to infer parameters for desired properties.

The currently available text is abstract-level information without figures or full methodological detail, so the specific form of model input features, the size of the library, the validation-set split, and the extent to which property-guided design was validated remain open questions that readers can confirm against the original figures and tables.