Designing chemical sensors for novel biomarkers is vital for future of wearable health monitoring devices. Polymer imprinted sensors have become a popular option due to their inherent selective nature but face complex design challenges. A polymer made 3-5 base units (monomers, co-monomers, crosslinkers, etc.) can have thousands of potential chemical formulations. Computational methods have sought to reduce experimental burden by simulating polymer–target interactions, most commonly using density functional theory (DFT). However, the associated computational costs limit discovery to only tens of potential polymers. Two classes of machine learning models have emerged to address this, those using raw 3D atomic coordinate inputs (e.g., SchNet, ANI), which are fast but less accurate, and those relying on semiempirical quantum inputs (e.g., OrbNet), which are more accurate but slower. Here, we introduce a deep learning framework that bridges this gap, achieving the accuracy of semiempirical models while matching the speed of 3D-coordinate-based approaches. This enables high-throughput evaluation of polymer–target interactions for chemical sensor development. Our model shows strong agreement with DFT-calibrated energies and exceptional scalability. Beyond biosensing, this platform can be applied to other areas where selectivity and design space are key, such as drug delivery and environmental toxin capture.
