Spatial transcriptomics (ST) has shown great potential for unraveling the molecular mechanisms of neurodegenerative diseases. However, most existing analyses of ST data focus on bulk or single-cell resolution, overlooking subcellular compartments such as synapses, which are fundamental structures of the brain's neural network. Here we present mcDETECT, a novel framework that integrates machine learning algorithms and in situ ST (iST) with targeted gene panels to study synapses. mcDETECT identifies individual synapses based on the aggregation of synaptic mRNAs in three-dimensional (3D) space, allowing for the construction of single-synapse spatial transcriptome profiles. By benchmarking the synapse density measured by volume electron microscopy and genetic labeling, we demonstrate that mcDETECT can faithfully and accurately recover the spatial location of single synapses using iST data from multiple platforms, including Xenium, Xenium 5K, MERSCOPE, and CosMx. Based on the subsequent transcriptome profiling, we further stratify total synapses into various subtypes and explore their pathogenic dysregulation associated with Alzheimer's disease (AD) progression, which provides potential targets for synapse-specific therapies in AD progression.