Abeer Fahad Abdullah Albakran

تخصص المعلوماتية الحيوية · دفعة 2026

Evaluating Binding Predictors for Aptamer Discovery: Construction Gaps in Interaction Benchmarks

Computational tools that predict whether an aptamer binds a target protein could accelerate aptamer discovery by shortlisting candidates before experimental testing. To be useful, predictors must work on proteins they have never seen. In this project, we investigated AptaTrans, a widely cited binding predictor reporting a ROC-AUC of 0.921. We re-curated its dataset and evaluated it under different ways of separating training and test data, including a setup where the model encountered entirely new protein targets during testing. We found that its performance depended strongly on whether the target protein had appeared during training, with performance on new targets close to chance. We also investigated how the training data and protein representations affected these results. Overall, the project highlights the difference between ranking aptamers for familiar targets and discovering binders for new ones, while providing a re-curated dataset and evaluation protocol for more rigorous testing.

اقرأ التقرير النهائي (PDF)