Morphology-guided convolutional Graph Neural Network decodes optically barcoded nanoparticles for one-pot, purification-, amplification-, and enzyme-free femtomolar nucleic-acid diagnostics

Affordable, accurate, and rapid point-of-care diagnostics remain elusive because existing nucleic acid tests generally require amplification, enzymes, purification, or prohibitively expensive equipment. Here, we present a one-pot assay that eliminates such requirements by pairing silver and gold nanoparticles with distinct light-scattering properties as optical barcodes, forcing hetero-particle pairing and allowing target-induced clustering to be imaged on a low-cost dark-field microscope, without costly fluorescent labels or bulky instrumentation. A novel deep learning architecture, Morphology-guided convolutional Graph Neural Network (Mc-GNN), combines morphology-guided kernel convolution and graph-based context-aware relational modelling to feature extract intuitively on single unbound particles and small cluster tensors, processing concurrently across ∼5000 particles per field using only <5 GB consumer GPUs. Mc-GNN, an ablation-validated novelty, near-perfectly classifies images by concentration down to femtomolar levels, with 98.2% average recall for the detection of synthetic DNA and performs robustly (94.8%) for RNA detection from whole SARS-CoV-2 virus, even with variations in nanoparticle selection and sample complexity, surpassing all baselines and popular architectures (≤ 90% for viral RNA). This computationally efficient, smartphone-compatible approach is readily extensible to new analytes and multiplexing, offering a scalable solution for fieldable diagnostics and future applications.