SIMPLEX: composition-space deep learning for bio-inspired hydrogel adhesion with prospective validation on model-discovered formulations and interpretable composition rules
Abstract
Fast and accurate composition-to-property prediction is critical for accelerating the discovery of multicomponent bio-adhesive hydrogels, yet current surrogate models are predominantly evaluated under random data splits that do not reflect the distribution shift encountered in real-world iterative screening. Moreover, existing deep learning approaches for small-sample formulation data offer little systematic guidance on architecture design and often depend on laboriously engineered interaction features. This study introduces Simplex-Inspired Multimodal Predictive Learning Encoder with X-modal Fusion (SIMPLEX), a dual-modality deep encoder purpose-built for composition-driven property estimation. Its gated multimodal fusion component autonomously learns inter-component interactions without hand-crafted quadratic terms. We establish a fully reproducible benchmark covering eight models with unified hyper-parameter tuning and statistical protocols. Evaluated on 316 internal hydrogel formulations under grouped cross-validation, SIMPLEX achieves first-tier predictive performance, competitive with the strongest classical baselines. On 25 prospectively collected formulations acquired from a Bayesian-optimized closed-loop campaign, a setting that introduces a deliberate distribution shift toward high-performance compositions, SIMPLEX delivers the highest coefficient of determination ( R 2 = 0.6712 ) for adhesion-strength prediction, meeting the demand for precise quantitative estimation in iterative candidate screening. Feature dissection reveals that the native compositional representation already captures the dominant composition–property relationships, and component-wise ablation pinpoints gated multimodal fusion as the principal performance driver. This study delivers a capable deep encoder together with a rigorous, transparent evaluation paradigm for formulation-driven machine learning, providing immediately reusable baselines and design principles for the community.