Artificial Intelligence-Driven SELEX Design of Aptamer Panels for Urinary Multi-Biomarker Detection in Prostate...

Background/Objectives: The limited specificity of prostate-specific antigen (PSA) drives unnecessary biopsies in prostate cancer (PCa). Urinary extracellular vesicles (uEVs) provide a non-invasive reservoir of tumor-derived nucleic acids and proteins. Aptamers selected by SELEX enable highly specific capture, and artificial intelligence (AI) can accelerate their optimization. This systematic review evaluated AI-assisted SELEX for urine-derived and exosome-enriched aptamer panels in PCa detection. Methods: Systematic searches of PubMed, Scopus, and Web of Science (1 January 2010-24 August 2025; no language restrictions) followed PRISMA 2020 and PRISMA-S. The protocol is registered on OSF (osf.io/b2y7u). After deduplication, 1348 records were screened; 129 studies met the eligibility criteria, including 34 (26.4%) integrating AI within SELEX or downstream refinement. Inclusion required at least one quantitative metric (dissociation constant K_d, SELEX cycles, limit of detection [LoD], sensitivity, specificity, or AUC). Risk of bias was appraised with QUADAS-2 (diagnostic accuracy studies) and PROBAST (prediction/machine learning models). Results: AI-assisted SELEX workflows reduced laboratory enrichment cycles from conventional 12-15 to 5-7 (≈40-55% relative reduction) and reported K_d values spanning low picomolar to upper nanomolar ranges; heterogeneity and inconsistent comparators precluded pooled estimates. Multiplex urinary panels (e.g., PCA3, TMPRSS2:ERG, miR-21, miR-375, EN2) yielded single-study AUCs between 0.70 and 0.92 with sensitivities up to 95% and specificities up to 88%; incomplete 2 × 2 contingency reporting prevented bivariate meta-analysis. LoD reporting was sparse and non-standardized despite several ultralow claims (attomolar to low femtomolar) on nanomaterial-enhanced platforms. Pre-analytical variability and absent threshold prespecification contributed to high or unclear risk (QUADAS-2). PROBAST frequently indicated high risk in participants and analysis domains. Across the included studies, lower K_d and reduced LoD improved analytical detectability; however, clinical specificity and AUC were predominantly shaped by pre-analytical control (matrix; post-DRE vs. spontaneous urine) and prespecified thresholds, so engineering gains did not consistently translate into higher diagnostic accuracy. Conclusions: AI-assisted SELEX is a promising strategy for accelerating high-affinity aptamer discovery and assembling multiplex urinary panels for PCa, but current evidence is early phase, heterogeneous, and largely single-center. Priorities include standardized uEV processing, complete 2 × 2 diagnostic reporting, multicenter external validation, calibration and decision impact analyses, and harmonized LoD and K_d reporting frameworks.