Machine learning-guided discovery of a bioactive hydrogel for spatiotemporally orchestrated tendon-to-bone healing.

Abstract

The rational design of biomaterials for complex tissue regeneration, such as the tendon-to-bone interface (TBI), is hindered by an immense combinatorial space that makes empirical optimization impractical. To address this challenge, a machine learning (ML) framework was developed to navigate this landscape. Leveraging a high-fidelity Tabular Pre-trained Transformer (TabPFN) model, a high-throughput virtual screen of over 137 million candidates identifies an optimal injectable microgranule-assembled macroporous hydrogel (MMH). The hydrogel is engineered to spatiotemporally deliver macrophage (M<i>ϕ</i>)-targeting small extracellular vesicles (CRV-sEVs) alongside Kartogenin (KGN)-loaded PLGA microspheres. In the initial phase, the release of CRV-sEVs reprograms the immune microenvironment toward a pro-reparative M2 phenotype by suppressing NF-κB signaling in M1 M<i>ϕ</i>. Subsequently, the sustained KGN release promotes endogenous stem cell chondrogenesis and fibrocartilage reconstruction. In a rat rotator cuff repair model, the ML-optimized hydrogel restores the structural integrity, biomechanical strength, and functional performance of the TBI, supporting the practical utility of the ML-guided discovery framework. This work offers a promising therapeutic strategy for TBI healing and establishes a data-driven paradigm for accelerating the discovery of complex multifunctional biomaterials for broad regenerative applications.