Reproduce “G2Φnet: Relating Genotype and Biomechanical Phenotype of Tissues with Deep Learning” on OpenResearch

About this paper

Enrui Zhang, Bart Spronck, Jay D. Humphrey, George Em Karniadakis

G2Φnet is a neural-operator framework that links genetic class and specimen-specific variation to the nonlinear biaxial mechanics of mouse aortas. Using sparse stress–stretch measurements, it reconstructs full constitutive responses with roughly 3.7–4.0% error and classifies four genotypes with up to 87% accuracy, though performance declines when genotypes have similar mechanical phenotypes.

Key insights

• A small set of measurements can recover a broad stress surface when the sampled points cover the deformation domain, with nine structured points yielding approximately 1.6–2.2% reconstruction error for a representative specimen.
• Classification and reconstruction favor different latent dimensions: greater flexibility improves fitting, whereas compact latent spaces generalize better, especially for extrapolation from unstructured measurements.
• The learned latent sample representation captures mathematical variation rather than a directly interpretable biological mechanism, and classification depends on genotypes producing distinguishable mechanical phenotypes.

Summary by alphaXiv.