Engineering an integrated biosensing interface combining DNA-assisted clustering and explainable AI for biomarker detection.

Point-of-care testing (POCT) platforms frequently suffer from a fundamental bottleneck: while advances in molecular amplification improve signal intensity, the reliability of signal readout in complex clinical matrices remains poorly controlled. Here, we present an integrated biosensing framework that treats readout reliability as an explicit engineering objective rather than a post hoc correction problem. The platform integrates three complementary components: (i) a heptameric nanobody probe employed as a multivalent recognition element for target capture, (ii) a DNA-assisted clustering interface that spatially organizes gold nanoparticle reporters for robust signal amplification, and (iii) a few-shot learning module based on Prototypical Networks that enables robust classification with minimal training data while providing interpretable decision-making through metric-based reasoning. Alpha-fetoprotein was selected as the model analyte because it remains a clinically important biomarker for hepatocellular carcinoma screening and follow-up, while also representing a realistic POCT challenge in which clinically meaningful detection must be achieved with low instrumentation burden and reliable readout under matrix variability. In this setting, the system achieves a visual limit of detection of 2 ng/mL and demonstrates quantitative consistency across representative clinical serum samples. Importantly, the AI module functions as an integral system component, identifying diagnostically relevant regions and mitigating readout uncertainty arising from matrix effects and imaging variability. By jointly engineering the sensing interface and the interpretive layer, this work establishes a generalizable strategy for constructing trustworthy POCT systems in which chemical signal generation and digital interpretation are co-designed.