Using TabPFN to simplify bioinformatics predictors: A case study using IApred Antigenicity Prediction

Background: Creating and fine-tuning computational predictors typically requires extensive hyperparameter optimization, particularly with traditional machine learning algorithms like Support Vector Machines (SVM). Tabular Prior-Data Fitted Networks (TabPFN), a transformer-based foundation model, eliminates hyperparameter tuning through in-context learning. Methods: We have recently developed IApred, an intrinsic antigenicity predictor using SVM. Now, we generated TabPFN models using the same training dataset and compared both methods' complexity and performance. While IApred-SVM required feature selection (k=529), 5-fold cross-validation hyperparameter tuning (RBF kernel, C=1, γ=0.001), and SMOTE preprocessing, IApred-TabPFN used the same data with minimal preprocessing (only VarianceThreshold to remove constant features), no hyperparameter tuning, and no SMOTE. Three TabPFN variants were evaluated: using all features, 529 features (matching SVM), and 100 features. Results: Comprehensive external validation on 436 independent proteins (218 antigens, 218 non-antigens) demonstrated consistent superiority of IApred-TabPFN variants over the extensively optimized IApred-SVM baseline. The TabPFN model using all features achieved the highest performance with ROC-AUC of 0.816 and MCC of 0.518, representing 4.3% and 20.4% improvements over SVM (ROC-AUC 0.782, MCC 0.430). Conclusions: TabPFN achieves superior predictive performance to optimized SVM while eliminating development complexity, representing a substantial reduction in development time without compromising accuracy.