VARDON: a deep learning framework with variational adaptive real dropout for enzyme classification in fish proteomes
Аннотация
Motivation Predicting enzyme function in non-model fish species remains challenging because of scarce experimental annotations and poor cross-species generalization of existing methods. Deep learning approaches using protein embeddings offer an efficient alternative, yet existing dropout mechanisms lack adaptability to embedding spaces and often fail to generalize across species. Results We propose VARDON (Variational Adaptive Real Dropout), a family of five novel dropout mechanisms with learnable per-neuron rates and Bayesian approximation. Using pre-computed UniProt embeddings (1024-dimensional) from 4,568 manually curated proteins across 12 fish species, we evaluated VARDON configurations against five baselines using 10-fold cross-validation, independent test sets, and leave-one-species-out (LOSO) cross-validation. VARDON variants demonstrated competitive performance on standard benchmarks (MCC = 0.8606, only 0.0008 behind the best baseline), while achieving superior cross-species generalization with VARDON_AdaptiveVD attaining the highest mean LOSO MCC (0.8823), essentially tied with the top-performing baseline (Residual_MLP: 0.8824) and outperforming Gaussian_Dropout_NN (0.8556) by 3.1%. VARDON variants also exhibited exceptional feature stability (up to 4.5 × higher than Gaussian Dropout), and statistical analysis using Friedman’s ANOVA revealed highly significant differences between methods ( p < 0.001), with VARDON variants winning 66.7% of configurations and 100% of AUC comparisons. Learnable dropout rates adapted to protein embedding spaces improved performance while reducing overfitting. Conclusions The proposed VARDON framework achieves competitive performance on standard benchmarks with superior cross-species generalization capability. The curated dataset of 4,568 reviewed proteins across 12 fish species and the source code are publicly available for reproducible research. Graphical Abstract