Machine learning-assisted washing-free detection of extracellular vesicles by target recycling amplification based fluorescent aptasensor for accurate diagnosis of gastric cancer.

Extracellular vesicles (EVs) are promising non-invasive biomarkers for cancer diagnosis. EVs proteins play a critical role in tumor progress and metastasis. However, accurately and reliably diagnosing cancers is greatly limited by single protein marker on EVs. Here, we reported an accurate diagnosis model of gastric cancer by analyzing five types of EVs surface proteins using machine learning in a retrospective study design. A washing-free detection method based on aptasensor and exonuclease Ⅰ was used to profile EVs surface proteins. The aptamer was designed as hairpin structure. The presence of target protein positive EVs converted the conformation of hairpin probes, which subsequently degraded by exonuclease Ⅰ. The exposed target protein could bind with and then open new hairpin probes, thus forming an amplification cycle. The lengths of different detection probes were optimized for detection. With the combination of five target proteins, five machine learning algorithms were compared to achieve a higher diagnostic accuracy. The best model, XGBoost, validated with 20 % of detection results could reach an accuracy of 0.8421. Furthermore, the XGBoost-based surface protein analysis could precisely identify gastric cancer patients with the area under the curve value of 0.9347 (95 % confidential interval (CI) = 0.8590 to 1.000). Since our method utilized a simple and versatile design of detection probes, its diagnostic scope could potentially be expanded to include different protein markers and accurately diagnose other diseases in the future.